A Regional assessment on the influence of climate change on summer rainfall : an application to shallow landsliding in Wanzhou County, China

A Regional Assessment on the Influence of Climate Change on Summer Rainfall: An application to shallow landsliding in Wanzhou County, China Final Thesis developed by: Joaquin Vicente Consunji Ferrer Directed by: Vicente Cesar De Medina Iglesia Carolina Puig Polo Marcel Hürlimann Master of Science in Flood Risk Management Barcelona, August 2021 Erasmus Mundus Programme in Flood Risk Management A Regional Assessment on the Influence of Climate  Change on Summer Rainfall:   An application to shallow landsliding in Wanzhou County, China    Joaquin Vicente Consunji Ferrer   Master of Science Thesis  August 2021  A Regional Assessment on the Influence of  Climate Change on Summer Rainfall:   An application to shallow landsliding in Wanzhou County, China    Master of Science Thesis  by   Joaquin Vicente Consunji Ferrer   Supervisors:  Vicente Cesar De Medina Iglesias (UPC Barcelona)  Carolina Puig Polo (UPC Barcelona)  Marcel Hürlimann (UPC Barcelona)   Examination committee  Vicente Cesar De Medina Iglesias (UPC Barcelona)  Allen Bateman Pinzon (UPC Barcelona)  Raul Sosa Perez (UPC Barcelona)     This research is done for the partial fulfilment of requirements for the Master of Science degree in the Erasmus  Mundus Flood Risk Management Programme.   Barcelona  August 2021   ©2021 by Joaquin Vicente Consunji Ferrer. All rights reserved. No part of this publication or the information contained herein may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, by photocopying, recording or otherwise, without the prior permission of the author. Although the author and institutions involved have made every effort to ensure that the information in this thesis was correct at press time, the author and institutions involved do not assume and hereby disclaim any liability to any party for any loss, damage, or disruption caused by errors or omissions, whether such errors or omissions result from negligence, accident, or any other cause. This work is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License. 0 Abstract Fatal landslides are devastating and widespread geohazard events that have a�ected millions of people and claimed the lives of thousands around the globe. Rainfall-induced landslides can occur at scales and magnitudes resulting in catastrophic damage and fatalities. Changes in climate have signi�cantly changed rainfall patterns, manifesting in more frequent extreme rainfall events around the world. The results of which have led to an increase in fatal landslides triggered by rainfall that is aggravated by human activity along sloped areas. Therefore, integrating climate projections with landslide susceptibility models will be crucial in assessing risk in the future. This study looks into the future of evolving triggering rainfall conditions in a changing climate, and aims to establish a link to tools in the present to assess shallow landslide susceptibility. Focus is given to the triggering rainfall conditions represented by extreme daily rainfall and mean seasonal rainfall. The study site selected in this research was Wanzhou County, China. This county lies in a region of China that receives 90% of its annual rainfall during the summer months. The e�ect of which is observed with 80% of shallow landslides occurring between June to August from 1995-2005. A methodology was developed to determine the triggering rainfall conditions from historical landslides, assess the rainfall conditions in the present, and bias-correct climate model outputs to obtain future projections for Wanzhou County. The results delivered in this research project provides practical value in integrating the climate change projections as inputs to physically-based shallow landslide susceptibility model inputs. This study �nds that the historical 30-day antecedent rainfall conditions that triggered shallow landslides during the summers of 1995 to 2005 were statistically similar to the mean seasonal rainfall derived from May to August, over the same period of time. This indicates that the mean seasonal rainfall can adequately represent the antecedent rainfall that can trigger shallow landslides over the summer season. The analysis of the present extreme daily rainfall �nds that a large central area of Wanzhou along the Yangtze River is currently exposed to high magnitude, low frequency events. The spatial distribution of extreme daily rainfall patterns suggests that the compounding events due to shallow landslides, urban �oods or �ash �oods along the Yangtze River should be thoroughly investigated. In order to assess climate change, an ensemble of four Regional climate model (RCM) outputs was considered. The RCM outputs were bias-corrected based on the daily distribution of precipitation observed between 1980-2018. The trends for extreme 1 i daily rainfall and mean seasonal rainfall were corrected independently. The climate change analysis derived future Climate Change Factors (CCFs) for scenarios in the Mid 21st Century (2021-2060) and the Late 21st Century (2061-2100). The climate change projections show a decrease in magnitude of extreme daily rainfall in the Mid 21st century. The Late 21 Century ensemble projections indicate a similar increases in magnitude for extreme daily rainfall and mean seasonal rainfall. The coe�cient of variation of the ensemble reveals less uncertainty in the ensemble projections for extreme daily rainfall compared to that of the mean daily rainfall. This research project delivered proof-of-concept for a methodological framework to derive shallow landslide triggering rainfall scenarios from climate model outputs. The presentation of the results and the identi�cation of sources of uncertainties in this study demonstrated a viable link between for climate change projections to provide future rainfall scenarios as inputs for physically-based shallow landslide susceptibility models. 2 ii 0 Acknowledgments This research would not have been possible without the supervisory committee at the Universitat Politècnica de Catalunya. I would like to extend my deepest gratitude to Vicente Medina for the brilliant scienti�c discussions and the deeply insightful life conversations over our cups of co�ee. I am extremely grateful to Marcel Hürlimann for trusting in my ability to deliver such extensive work in your research group, and guiding me during our weekly scienti�c discussions. I am thankful to Carol Puig for directing my thesis away from the infeasible, and rekindling my excitement for �eld work and travel. I would like to extend my sincere thanks to Zizheng Guo for the trust to support your doctoral research. I also wish to thank Ona Torra and Claudia Abanco for sharing your studies and insights to improve my research project. I am thankful to Schalk Jan van Andel for the insight in �nalizing this thesis. I gratefully acknowledge the opportunity provided by the Erasmus Mundus Flood Risk Management Programme consortium and the Erasmus+ Scholarship to pursue my graduate studies. I am grateful to all the professors and lecturers in the consortium who have dedicated time and e�ort to provide meaningful education throughout this pandemic. I’m thankful for the many life lessons I’ve learned from my FRM8 batchmates. I’m grateful to Marcos, Eugen, Manuel, Rodrigo, and Siamak; thank you for your friendship and all our adventures during this pandemic. A special thanks to Dominique, Analia, Siva, Faisal, Tharindu, Amin, Helen, Nele, Alina, Camilo, Valeria, Manoel, and Yue for the debates and discussions during our case studies that have expanded my perspective. Thank you, Buse; you were always there to support, and motivate me. I’m deeply thankful to my family in Galizano for being home away from home. I owe my deepest gratitude to my mom, dad, and brother. Thank you for the endless encouragement, love and support. You are my pillars. This masters research was preformed in the middle of the COVID-19 pandemic. I owe a debt of gratitude to all those who have supported me in completing this program. 3 iii 0 Contents Abstract 1 Acknowledgments 3 Contents 5 1 Introduction 1 1.1 Background ................................. 1 1.2 Formulation of Research Questions .................... 1 1.3 Study Area .................................. 6 1.4 Research Objectives ............................. 7 1.5 Innovation and Practical Value ....................... 8 2 State of the Art 9 2.1 De�nition of Terms ............................. 9 2.1.1 Triggering Rainfall Conditions .................. 9 2.1.2 Extreme and Mean Daily Rainfall Conditions .......... 10 2.2 Determination of Triggering Rainfall Conditions ............. 12 2.3 Shallow Landslide Susceptibility and Climate Change .......... 16 3 Case Study on Wanzhou County, China 19 3.1 General Site Description .......................... 19 3.2 Landslide Studies in Wanzhou County ................... 20 3.3 Shallow Landslides in Wanzhou ...................... 24 3.4 Climate Impact Assessments on the Upper Yangtze River Basin ..... 26 4 Methodology 33 4.1 Methodological Framework ......................... 33 4.2 Precipitation Data and Climate Model Outputs .............. 33 4.2.1 Gridded Precipitation Observations ................ 33 4.2.2 Multi-model Climate Projection Ensemble ............ 36 4.3 Reconstruction of Triggering Rainfall Conditions ............ 39 4.4 Extreme Daily Rainfall Analysis ...................... 41 4.5 Climate Change Analysis .......................... 44 4.5.1 Bias Correction Methodology ................... 44 5 v 5.40 Coe�cients of variation for Mid 21st Century scenario projection ensemble projections of climate change factors ( ⇠⇠⇢⇡',) ) with return periods ( ) )of 5, 10, 50 and 100 years derived from the ensemble mean projections covering the months of June to July. .............112 5.41 Coe�cients of variation for Late 21st Century scenario projection ensemble projections of climate change factors ( ⇠⇠⇢⇡',) ) with return periods ) of 5, 10, 50 and 100 years derived from the ensemble mean projections covering the months of June to July. .............113 12 xii 0 List of Tables 4.1 Summary of ensemble speci�cations. ................... 38 4.2 De�nitions of climate change scenarios. .................. 45 5.1 Summary of the Kolmogorov-Smirnov (KS) Test results between CDFs of the original record date data and the 7-day time-shifted data for di�erent antecedent durations. ....................... 75 5.2 Summary of minimum, mean and maximum Gumbel �t parameters. .. 94 5.3 Summary Goodness-of-�t test statistics with p-values (in parenthesis). 94 5.4 Summary of the Kolmogorov-Smirnov (KS) test between bias-corrected ensemble members with minimum, mean and maximum test statistics and p-values (in parenthesis). ........................102 5.5 Summary of the bias-corrected RCM cells passing a Kolmogorov-Smirnov Test and the total number of cells corrected with a ?{0;D4 0.05 ..104 13 xiii 0 List of Abbreviations AD Anderson-Darling AHI Advanced Himawari Imager APHRODITE Asian Precipitation—Highly Resolved Observational Data Integration Towards the Evaluation of Water Resources CAR Calibrated Antecedent Rainfall CCF Climate Change Factor CDF Cumulative Distribution Function CMA Chinese Meteorological Agency CMFD China Meteorological Forcing Dataset CMIP5 Coupled Model Intercomparison Project 5 CORDEX Coordinated Downscaling Experiment CVM Cramer-von-Mises EDR Extreme Daily Rainfall EVS Extreme Value Statistics FSLAM Fast Shallow Landslide Assessment Model GCM Global Climate Model GERICS Climate Service Center Germany GEV General Extreme Value GLDAS Global Land Assimilation System GPM Global Precipitation Measurement HadGEM Hadley Global Environment Model KS Kolmogorov-Smirnov MAE Mean Absolute Error MMR Mean Monthly Rainfall 15 xv Chapter 0 List of Abbreviations MSR Mean Seasonal Rainfall MPI Max Planck Institute MPI-ESM Max Planck Institute Earth System Model NDVI Normalized Di�erence Vegetation Index RainFARM Rainfall Downscaling by a Filtered Auto-Regressive Model RCM Regional Climate Model RCP Representative Concentration Pathways REA Reliable Ensemble Averaging RegCM Regional Climate Model of the International Centre for Theoretical Physics’ REMO Regional Climate Model of the Climate Service Center Germany RMSE Root-mean-squared Error QDM Quanntile Delta Mapping QM Quantile Mapping TMPA TRMM Multi-satellite Precipitation Analysis TRMM Tropical Rainfall Measuring Mission TRIGRS Transient Rainfall In�ltration and Grid-Based Regional Slope-Stability Model UYRB Upper Yangtze River Basin YRB Yangtze River Basin WRF Weather Research and Forecasting 16 xv i 1Introduction 1.1 Background Fatal landslides are devastating and widespread geohazard events that have a�ected millions of people and claimed the lives of thousands around the globe [1]. Froude & Petley [2] have determined that rainfall is the main driver of fatal landslide occurrences between 2004-2016. Furthermore, it was determined that landslides triggered by human activity along the slopes is increasing. As society looks to manage landslide risk in the future, consideration of the causal relationship between landslides and climate change is essential [3]. Therefore, the interaction between rainfall under a changing climate and the increase in human activity is critical to assessing and managing landslide risk in the future. There is evident impact of anthropogenic in�uences on the change in climate at a global scale throughout the Industrial Period from 1750-2011 [4]. Centuries of increasing widespread land cover changes driven by human activity have resulted in the reduction of global forest areas [5]. This trend of declining forest cover across the globe a�ects mountainous and sloped areas through the alteration of vegetation and soil stability characteristics. The changes in global climate are apparent in the observation of trends in precipitation and records of more frequent extreme precipitation events. An empirical assessment of the upper 0.3% of daily rainfall observations showed a widespread increase in frequency during the past 50 to 100 years [6]. Alexander et al. [7]�nd a period of accelerated global warming between 1964-2013, where 7% more extreme events were observed. The temporal variation in rainfall characteristics is a signi�cant driver in landslide initiation across the world. A strong correlation between mean monthly rainfall and landslides was observed in Central America, South America, South Asia and East Asia. Increased human activity through land development and illegal mining practices are foreseen to instigate landslides in the future [2]. 1.2 Formulation of Research �estions The relationship between climate change and landslide risk is a critical topic in regions across the globe. Gariano & Guzzetti [3]�nd a causal relationship between landslides 1 Chapter 1 Introduction and climate across literature. Although an increase in risk is expected in the future, quantifying risk to the population and projecting the precise impact of climate change is di�cult due to large uncertainty derived from the landslide-climate projection modelling chains. Peres & Cancelliere [8] observed signi�cant indicators of seasonality and interannual variation in the occurrence of landslides from 2004-2016. Their �ndings suggest that antecedent conditions in�uence soil moisture and interact with mechanical soil properties to increase the probability of slope failure. This research focuses on rainfall-induced shallow landslides that occur on sloped surfaces and are capable of causing devastating damage and massive fatalities. Li & Mo [9] provide a contemporary classi�cation of shallow landslides determined by the depth of the rupture surface being less than 10 meters, consisting of mostly soil material with possible soft and hard rock from the slip surface. Rainfall-triggered shallow landslides occur due to rainfall events that increase the top-soil water content and pore pressure. In e�ect, the e�ective stresses are reduced, the soil is destabilized and a debris �ow that can propagate a signi�cant volume of soil mass several kilometers at high velocities can be mobilized. This characteristic propagation of mass particular to the shallow landslide results in deadly debris �ows with heightened risk to communities along the �ow paths. In cases where soil on a slope with signi�cant soil moisture is subjected to heavy rainfall, a faster propagation of wetting between the soil-bedrock interface can create conditions of landslide initiation [10]. There are multiple approaches in modeling to quantitatively assess shallow landslide susceptibility. The classi�cation of these methods mainly use statistical models, machine learning models and physically-based models. Although data-driven models are reliable and widely accepted, they do not describe the complex interactions occurring and physical processes that govern the behaviour of shallow landslide susceptibility and thus risk in future scenarios [3,11,12]. Physically-based models are able to describe the processes of soil-vegetation interactions, in�uences of event rainfall on in�ltration, and the e�ects of antecedent rainfall events to subsurface �ows. These processes are essential to assess the shallow landslide risk in the future under the conditions of a changing climate and changes to land use. A gap in research on modeling shallow landslide susceptibility under climate change scenarios exists in the inability of current models to incorporate the non-stationary trends of increasing precipitation, and the e�ects of speci�c triggers on the magnitude of future landslides. [3] identi�ed these issues and recommended the development of new models to cope with the non-stationary climate characteristics, and investigate temporal variations in risk that are driven by a combination of climate and vegetation changes along the slope environments. Their research indicates the in�uence of antecedent rainfall, and land use change scenarios that induce stable slopes into marginally stable conditions, while bringing 2 Formulation of Research �estions Section 1.2 marginally stable slopes into critical risk conditions. Modeling climate change requires incorporating interactions between stability under antecedent rainfall events, and changes in vegetationderived resistance. All are essential when analyzing risks against the future impacts of landslide-triggering extreme precipitation events. Therefore, this research project aimed to address the gap in research when modeling shallow landslide susceptibility under climate change scenarios. A focus on evaluating future rainfall conditions that can trigger shallow landslides was the primary research objective with the question stated as: What is the influence of climate change on rainfall conditions that could trigger shallow landslides? In order for this research project to accomplish its main objective, a set of auxiliary objectives were established. The �rst auxiliary research question revolves around understanding the rainfall conditions that triggered previous shallow landslide events. Physically-based models require scenarios of triggering rainfall conditions that incorporate the non-stationary trends of increasing precipitation to determining areas at risk under climate change conditions. This study particularly supports the development of rainfall scenarios that are compatible with the The “Fast Shallow Landslide Assessment Model” (FSLAM), developed by [11]. This model utilizes geotechnical and hydrological models to enable the investigation of future climate and land use scenarios involving changes in daily rainfall patterns and soil cohesion. The geotechnical model incorporates cohesion derived from root strength alongside cohesion in the soil matrix to account for vegetation characteristics accompanying environmental changes. Furthermore, the hydrological model combines the in�uences of antecedent recharge through a steady-state lateral �ow, and event precipitation through vertical �ow on the water table. While the sensitivity analysis of the model reveals antecedent recharge has a considerable in�uence over the probability of failure compared to the event rainfall. A fundamental variable in the performance of models is the antecedent rainfall. It is recognized in literature as an in�uential factor in the occurrence of soil slips and thus, the initiation of shallow landslides [13]. In the determination of rainfall thresholds for landslide occurrences, approaches with antecedent rainfall conditions comprise the second most used approach. A variety of methods in incorporating antecedent rainfall have been observed, where some studies prefer the use of rainfall indices instead of direct rainfall measurements to account for the in�ltration, storage and saturation [14]. The use of indices to represent antecedent rainfall and account for soil storage capacity is an important consideration of the dynamics of soil moisture during event intervals [10]. It is di�cult to generalize the approach and threshold to which antecedent rainfall in�uences shallow landslide initiation due to the regional climate interaction with in situ geotechnical properties and watershed’s hydrological characteristics [13]. 3 Chapter 1 Introduction The importance of antecedent rainfall as a factor for rainfalltriggered landslides was observed in Calabria, Italy, by Polemio & Petrucci [15]. This research project identi�ed shallow landslide triggering rainfall conditions as event rainfall and antecedent rainfall. The �rst auxiliary research question was formulated to evaluate the rainfall conditions in that past that had led to landslide occurrences in survey inventories and literature. This auxiliary question is stated as: What rainfall conditions triggered the historic shallow landslides in the study area? This research question was intended to develop a task to use inventory data to reconstruct the event rainfall, antecedent rainfall, and antecedent recharge in order to derive insight on patterns and trends in landslide triggering at a site-speci�c scale. The understanding of the rainfall conditions that triggered landslides in the inventory aid to establish a reference to determine rainfall conditions in the present that could trigger landslides in the near future. Thus, an assessment of the present conditions is necessary to capture the recent spatio-temporal patterns. The �ndings of Alexander et al. [7] indicate that while there is signi�cant warming throughout the 20th century and a general increase in precipitation volume from 19512003 was detected at a global level. Donat et al. [16] presented an analysis of “wet” and “dry” regions determined by changes in total precipitation from global observation data sets for the period of 1951–1980. Their study showed a statistically signi�cant increase in annual-maximum daily precipitation observed in both wet and dry regions around the globe. The precipitation observations indicate that the present climate conditions are characterized by the dynamics of a changing climate that have already seasonal volumes of precipitation, and maximum daily precipitation [4]. This insight is critical to the shallow landslide triggering rainfall conditions. An increase in precipitation volume at a seasonal level a�ects antecedent rainfall conditions, while a trend of increasing maximum daily precipitation could in�uence the frequency of event rainfall. Furthermore, the importance of incorporating the accompanying frequency of extreme events is emphasized by Scheidl et al. [12] in a study that simulated the interaction of changing vegetation cover with design precipitation events with 100-year recurrence. Polemio & Petrucci [15]�nd that the number landslides was correlated to monthly precipitation, wet days and precipitation intensity. Therefore, a site-speci�c analysis of the prevailing rainfall conditions is essential to establish a reference scenario for future rainfall projections in a changing climate. An auxiliary research question developed to establish a reference scenario based on present rainfall conditions is stated as: What are the return periods of extreme daily rainfall and the magnitudes of seasonal rainfall? 4 Formulation of Research �estions Section 1.3 Gariano & Guzzetti [3] expect an increase in frequency and intensity of severe rainfall events, that will result increase in the number of people exposed to landslide risk. Scheidl et al. [12] indicated that a general increase in shallow landslide risk was observed under climate change, highlighting the signi�cance of the change in frequency of extreme precipitation events as a triggering mechanism. The incorporation non-stationary events under climate change will enhance the analysis for adaptive management approaches by accounting for the in�uence of the changing climate on the return periods of triggering event rainfall. Therefore, the �nal auxiliary research question was formulated to assess the change in frequency and magnitude of extreme daily rainfall, and the volume of seasonal precipitation in the future. This research question is formulated as: How do the distributions and magnitudes of rainfall events change based on climate models projections? The main goal of the research task following this auxiliary research question is to derive projections of climate change conditions from an ensemble of Regional climate models (RCMs) selected for the study area. An ensemble of three RCMs were utilized by Peres & Cancelliere [8] to investigate landslide-triggering return periods. Although there was an indication of projected increases in the interarrival time between rainfall events, the high spread of spatial variation stipulates the e�ect of model uncertainty and points to the use of a wider range of models in an ensemble to analyze the impacts of climate change. An analysis of annual maximum 1-day precipitation events under climate change scenarios by Brönnimann et al. [17] shows a shift in seasonality that results in less frequent events. Scheidl et al. [12] suggests that in regions where extreme precipitation events are less likely to occur under drying climate scenarios, a decrease in frequency of landslide-triggering events can be expected. Thus, indicating that there is signi�cant uncertainty and variability from assessing landslide risk using rainfall scenarios from projections of the numerous available regional climate models. Translating projected rainfall magnitudes as direct input to physically-based susceptibility models will introduce signi�cant uncertainty into the risk assessments. Therefore, the perspective in assessing the in�uence of climate change adapted to this research project manifested in the calculation of a factor of change derived from the climate model projections. This captured relative change in the climate signal that could be applied to the reference scenario in the present to project future events. 5 Chapter 2 State of the Art 2.2 Determination of Triggering Rainfall Conditions The role of antecedent rainfall as a determining factor for landslide initiation is recognized through the extensive studies on deriving statistically-based thresholds, investigating water table �eld observations and in in�uencing slope stability of physically-based models [23–31]. Relevant literature in de�ning antecedent rainfall events in triggering shallow landslides are discussed in this section. Segoni et al. [32] observes that while antecedent rainfall is widely used in establishing rainfall-duration thresholds, there is an overwhelming variety of approaches taken by research groups in treating these events. Although this is the case, it was observed that researchers often do not directly use rainfall measurements in incorporating antecedent rainfall into their studies, but rather derive an antecedent rainfall index to incorporate degrees of saturation across the study area’s terrain. The processing of rainfall measurements to reconstruct an antecedent rainfall index is performed to account for a decrease in in�uence of rain events with respect to time due to the drainage processes across the terrain. A calibrated antecedent rainfall index was proposed by Crozier [23] and is presented in in Equation 1: ⇠'G== %1+ 2%2+... + =%=(2.2) where ⇠'G is the calibrated antecedent rainfall index for day G ; %= is the daily rainfall G days prior to G . The constant is the decay constant and an empirical parameter dependent on the draining capacity and the hydrological characteristics of the area. In order to determine the -value decay constant, studies conduct trials but a consensus arrives at =0.9 with a limit to 30-day antecedent rainfall in study sites spread across Bangladesh, Korea, and Portugal [25,26,28]. The study of Kim et al. [26] assessed that an appropriate determination of a calibrated antecedent rainfall index to identify landslide-triggering events at a regional scale requires a detailed study on the hydrological processes, instead of insight from a datadriven analysis. The calibrated antecedent rainfall index is used to de�ne an e�ective antecedent rainfall-duration threshold to analyze patterns of rainfall-triggering rainfall conditions and determine thresholds for operational early warning systems. Although this is a widely used method, the results of Ma et al. [27] on a study area in Zhejiang, China, show that the false positive rates for -values between 0.9 and 0.75 go from 63% to 51%. While this method is useful for operational purposes and early warning systems in ungauged basins, the false-positive rates and spatial variation of the hydrological processes across a catchment should be considered when being used to de�ne an appropriate constant for identifying critical antecedent rainfall. Another method in determining the duration to which antecedent rainfall can trigger 12 Determination of Triggering Rainfall Conditions Section 2.2 a shallow landslide is a comparison of cumulative antecedent rainfall and event rainfall using a 1:1 line to graphically assess the correlation between antecedent rainfall and event rainfall as triggering rainfall conditions, as illustrated in Figure 2.3. A study by Dahal & Hasegawa [33] on landslides in the Nepalese Himalayas utilizes this method to determine the role of cumulative antecedent rainfall versus event rainfall, based on an inventory of historical landslides. The population bias is measured across a 45-degree line on a bivariate plot to determine the in�uence of either event rainfall or antecedent rainfall. It was determined that while 55% of inventory landslide events considering 3-day antecedent rainfall were determined by the event rainfall, 30-day antecedent rainfall determined 98.4% of landslide events when compared to the event rainfall. Furthermore, this study examines the relationship between slope failure, and daily rainfall versus cumulative rainfall in the Nepalese Himalayas during the critical monsoon season. The results using the same method �nd that the 10-day cumulative rainfall at failure showed a higher correlation coe�cient compared to 3,5, and 30-day periods. The results depicted in Figure 2.3 demonstrate insight that can be derived through the method of population bias measurement versus a 1:1 diagonal line on a bivariate plot and an analysis correlation coe�cient for di�erent antecedent event periods. Mathew et al. [29] utilized the same method of analysis in the Garhwal Himalaya, India, to assess combinations of rainfall parameters and their in�uence on triggering slope failures. Their analysis found that comparing daily rainfall values of the failure events to the 3-day cumulative rainfall and 30-day cumulative rainfall yield identical results. In this area, the distinction of antecedent rainfall in terms of in�uence on determining slope failure is not evident. The literature reviewed on studies utilizing the 1:1 bivariate plot method have shown how this simple approach to establishing a correlation between cumulative antecedent rainfall and event rainfall has been e�ective in analyzing the in�uence of antecedent rainfall, identi�cation of major rainfall-triggers for shallow landslides and the e�ects of di�erent durations in antecedent rainfall and event rainfall on landslide occurrence. The methods of antecedent rainfall index and the 1:1 bivariate plot analysis are data-driven insights to fortify the de�nition of antecedent rainfall in an area, region or study site. Observations from �eld experiments and insight from physically-based models account for responses to geotechnical and hydrological processes and provide more de�nitive thresholds for durations critical antecedent rainfall. Tu et al. [31] in a �eld experiment on an engineered loess slope, subjected the soil to arti�cial rainfall and measured the response of the slope with instruments to measure matric suction, volumetric water content, and pore pressure at di�erent depths. Their �ndings revealed that matric suction decreases rapidly for 120 mm/day rainfall, with an in�uence depth reaching 3 meters. It was determined in this study that the 9-day 13 Chapter 2 State of the Art Figure 2.3: Comparison of the relationship between cumulative rainfall before failure and daily rainfall at failure over the Nepal Himalayas for di�erent durations of consideration. The broken line (guide line) represents daily and cumulative rainfalls are the same at failure. Cited from Dahal & Hasegawa [33]. 14 Determination of Triggering Rainfall Conditions Section 2.2 antecedent rainfall was critical to slope stability, where the matric suction at a 3-meter depth was almost zero. The analysis of results from this study suggests that it is possible for 9-day antecedent rainfall with 120 mm/day event rainfalls to have serious consequences on slope stability for loess-type engineered slopes. While this may not be directly applicable to all slopes, these results provide a reasonable starting point for further analysis with a data-driven approach. Tang et al. [30] used a physically-based model using two-dimensional seepage analysis to analyze the e�ects of antecedent rainfall with data from the Three Gorges Reservoir Area, China, with a homogeneous slope model. The model results show signi�cant variation depending on the permeability coe�cient of the soil, thus deriving di�erent conclusions for sandy and clayey soil type slopes. In particular, the recommendations look at minimum durations of consideration following heavy rainfall events, and maximum durations, where the in�uence of antecedent rainfall is minimal. It was concluded that the 15-day antecedent rainfall was the minimum to re�ect a real factor of safety for slope stability for sandy slopes, while the in�uence beyond a 30-day antecedent rainfall duration is minimal. The results for clayey slopes suggest a minimum consideration of 20-day antecedent rainfall and a maximum of 40-day antecedent rainfall. While the model results show that the minimum factor of safety does not have a direct relationship with the maximum daily rainfall, the maximum accumulating rainfall over a 10-day period for sandy slopes and a 15-day period for clayey slopes can be related to a minimum factor of safety. The recommendations from this physically-based model expands the results from �eld observations to provide initial values in de�ning signi�cant antecedent rainfall duration. The analysis of literature in this section provides reasonable initial considerations of signi�cant antecedent rainfall duration based on �eld experiment observations as well as physically-based model analysis of slope stability for di�erent soil types. It was also established that two methods that can be used in assessing the agreement of the initial critical antecedent rainfall duration and the observed landslide occurrences from an inventory of historical events. 15 Chapter 2 State of the Art 2.3 Shallow Landslide Susceptibility and Climate Change This section analyzes studies and literature that attempts to estimate future landslide susceptibility with Global Climate Models (GCMs) or Regional Climate Models (RCMs), and looks at insight on means of dealing with the uncertainty and bias inherent to landslide-climate modeling chains. Gariano & Guzzetti [3] presented a framework for landslide-climate modeling and recommended that scenarios investigating the impact of climate and environmental changes on landslide susceptibility is crucial to understanding the future of landslide risk, shown in Figure 1.2. A current gap in the landslide-climate modeling chains are found in the lack of performance of global climate models, and the uncertainty from climate change scenarios. It was determined in their study that regional-scale study results are strongly dependent on the selection of the driving models, the downscaling methods and the adopted reference scenarios. These factors reduce con�dence in the landslide projections from studies, and a way forward can be found in the utilization of an ensemble of models to quantify the uncertainties of projections. Alvioli et al. [34] investigates the relationship between downscaled climate projections and landslide occurrences for the years 2010-2049. The study used a model chain consisting of Weather Research and Forecasting (WRF) RCM simulation outputs with Rainfall Downscaling by a Filtered AutoRegressive Model (RainFARM) to provide input rainfall data, while using the Transient Rainfall In�ltration and Grid-Based Regional Slope-Stability Model v. 2.0 (TRIGRS) slope stability model to analyze the future landslide susceptibility. The resulting increased uncertainty between the model evaluation based on measured events and the distribution of events from the output of downscaled RCM. The large uncertainties in the landslide-climate modeling chain were found to make this approach less robust than establishing rainfall thresholds. The parametrization, resolution and processes of di�erent climate models were established to be an important source of potential uncertainty in the conclusion by this study. Peres & Cancelliere [8] combined the use of a stochastic rainfall generator, a simpli�ed Richards vertical in�ltration hydrological model, and an in�nite slope model to estimate landslide probability through Monte Carlo simulations. This study investigated the impacts of climate change projections from an ensemble of three RCMs through the extraction of change factors to adjust duration and rainfall depth parameters in the stochastic rainfall generator. Their results �nd that the change factors in Representative concentration pathways (RCPs) 4.5 and 8.5 re�ect signi�cant seasonal variation, and the results indicate a variation between RCM results based driving GCM a�ects the projected frequency of rainfall events. The changes in standard deviation are found to be more signi�cant than the mean precipitation. The results of the study implementing this combined modeling framework indicate that the presence of a high spread of climate modeling uncertainty can be partially 16 Shallow Landslide Susceptibility and Climate Change Section 2.3 addressed with the use of a wider ensemble of RCMs. This study demonstrates a simpli�ed approach to utilizing projections from RCMs in establishing trends, and characteristics of climate change in reducing uncertainty along the landslide-climate modeling chain. Scheidl et al. [12] investigated combined climate change and land use change scenarios through an integrated landslide-climate-landscape modeling chain. Although climate scenarios were considered in this study for the landscape simulations, three design rainfall events were considered in this study based on a 100-year recurrence. The recommendations identi�ed incorporating possible in�uences derived from the change in frequency of extreme precipitation events under climate change in future studies. The review of literature in this section looks at results and recommendations from studies using landslide-climate model chains. It was observed that the greatest source of uncertainty in the model chain is from the climate models and the resulting projections [3,34]. It has been suggested that utilizing a wider ensemble of GCM-RCM combinations will allow for the quanti�cation of uncertainty in determining impact of climate change on shallow landslide risk [8]. Scheidl et al. [12] indicated that another approach for long-term modeling is the simpli�cation of representing triggering rainfall events through implementing design rainfalls with a return period. Careful consideration must be taken into account when specifying return period events in future cases, due to the changing distribution of events driven by the non-stationary nature of climate change. Finally, Peres & Cancelliere [8] present the idea of deriving change factors from global climate models instead of directly applying downscaled RCM results to shallow landslide models is an approach to reducing uncertainty in a complex modeling chain. 17 3Case Study on Wanzhou County, China A focus on the descriptive site details and contemporary landslide research is presented for the selected study area of Wanzhou County, China. Climate change impact research and assessments over China and speci�c to the Yangtze River Basin are also discussed. The discussion of this chapter gives context to the trends and factors that could a�ect the study site from a wider climate outlook. 3.1 General Site Description Wanzhou County, China was selected as a study site of this research project and accomplish the objectives presented in Section 1.4. The study site is bounded by the coordinates N 30 ° 24 0 25 00 , E 107 ° 52 0 22 00 to N 31 ° 14 0 58 00 , E 108 ° 53 0 25 00 , with an area of 3 457 km 2 . The district is located within Three-Gorges Reservoir along the Yangtze River, and upstream of the Three Gorges Dam, shown in Figure 1.1. This region of China is categorized as subtropical humid monsoon zone with a mean annual precipitation of 1191.3 mm. The precipitation patterns in this area are characterized by 90% of the annual rainfall occurring between May and September. Summer rainfall can be characterized by short intense rainstorms with daily rainfall values exceeding 100 mm/day [19]. The Lithology in the Wanzhou district comprises of the a Jurassic (J) system and a Triassic (T) system, illustrated in Figure 1.1. The Jurassic system mainly consists of mudstone, sandstone, siltstone and shale. The Triassic system is primarily composed of limestone, clay stone, sandstone, siltstone, and coal. 19 Chapter 3 Case Study on Wanzhou County, China Figure 3.1: Lithological units in Wanzhou district: 1 (quartz sandstone, shale, limestone), 2 (sandy mudstone, quartz sandstone, shale, feldspatite, siltstone), 3 (quartz sandstone, aubergine mudstone, lithic sandstone, shale), )1 (dolomitic limestone, �aglike limestone), )2 (argillaceous limestone, dysaerobic fauna, aubergine mudstone, calcareous shale), and )3 (lithic Sandstone, arenaceous shale, coal seam). Cited from Huang et al. [20]. 3.2 Landslide Studies in Wanzhou County Landslide occurrences in the Wanzhou County are of concern, with large and numerous events occurring over recent years. Such frequent occurrences have been a subject of interest for advances in research towards better landslide susceptibility mapping and disaster risk reduction. The heavy rainfall in the study area was attributed to be a driver of the signi�cant and frequent landslides that have occurred. [20,21,35]. In 2016, the Wanzhou District was selected to pilot community-based DRR program to reduce landslide risk instigated by survey �ndings indicating the occurrence 921 landslides surveyed in the region. It was determined that many of these landslides occurred after the �lling and drawdown operations of the Three Gorges Dam [18]. The Tangjiao, Sanzhouxi, and Jinzhuquji landslides, shown in Figure 3.2, are examples of landslides of concern that were monitored and studied in the implementation of this program. In another study, an inventory of 639 rock and soil landslides occurring between 1970 to 2013 was utilized to explore machine learning techniques for susceptibility mapping [20]. The analysis of this inventory �nds that nearly half of the 639 landslides occurred 20 Landslide Studies in Wanzhou County Section 3.2 Figure 3.2: Locations of the 639 distinct landslides inventoried within the Wanzhou district from 1970 to 2013 Cited from Huang et al. [20]. between 1993 to 2013, while the rest were reactivated old landslide sites. The triggering factors for soil landslides in this inventory were determined to have been triggered by heavy rainfall and in�uence from �uctuations in the reservoir level. Anthropogenic activities that involve slope modi�cation were attributed as an increasingly in�uential triggering factor. The susceptibility map derived from the implementation self-organizing-map network and extreme learning machine method in the study of Huang et al. [20], is presented in Figure 3.2. Their analysis of the landslide triggering mechanisms based on the inventory resulted in the development of the data-driven model around geophysical characteristics, land cover, and distance to the Yangtze River. Naturally, the critical areas within the district are present in proximity to the river. The �nal result of this study was a susceptibility map for Wanzhou County developed with self-organizing -map network and extreme learning machine. The implementation of data-driven approaches to perform landslide susceptibility mapping in Wanzhou were further developed in studies by Song et al. [35] and Xiao et al. [36] to implement a variety of statistical-based machine learning techniques. The test site of Song et al. [35] particularly focused on the river area, showing more detailed results than those in Figure 3.3 by incorporating annual rainfall as the most important factor in determining landslide susceptibility. 21 Chapter 3 Case Study on Wanzhou County, China Figure 3.9: Hydrological basins within the Upper Yangtze River Basin. Cited from Huang et al. [38]. study were driven by the GCMs submitted to the Coupled Model Intercomparison Project Phase 5 (CMIP5). The models were corrected for systematic model biases with the parametric quantile mapping method though a mixture of �tted frequency distributions. The �ndings for the near-future scenario project a warm-humid climate in the western UYRB with an expected decrease of precipitation at a rate of 19.05–19.25 mm/10 a. Furthermore, variation of precipitation on a multiyear average annual basis is projected to vary between -0.5 to 0.5 mm/day in the near-future precipitation projection in this study exhibits an insigni�cant downward trend, with signi�cant variation between east and west of the UYRB. The multiyear average precipitation in the Sichuan Basin signi�cantly increased under the RCP 8.5 climate change scenario in the northwestern region while signi�cantly decreasing in the southeastern areas. The separation of regions projecting increase and decrease across the Sichuan basin are illustrated in Figure 3.10. The discussion of this section gives a brief overview of the insight derived from climate change studies including the UYRB. It is evident that there are limitations in studies that derive insight from coarse resolution climate models in capturing the critical East Asian monsoon. This is a critical limitation for climate studies over the UYRB and in e�ect, for climate studies about Wanzhou County, since this region’s precipitation 28 Climate Impact Assessments on the Upper Yangtze River Basin Section 3.4 Figure 3.10: Near-future multi-year average changes in precipitation (mm/day) over the UYRB under the RCP8.5 scenarios compared to reference observations from 1971 to 2000. Cited from Huang et al. [38]. patterns are a�ected by the interaction of the East Asian and Indian summer monsoons. Thus, utilizing �ner resolution RCMs should be a consideration in capturing that climate dynamics that are critical to assessing climate change. The spatial distribution of change precipitation in the near-future, compared to the reference observations shown in Figure 3.10 places Wanzhou at the border of the region of signi�cant increase and decrease. The study area is approximately place within the region of a mild decrease in precipitation in the near-future under the RCP 8.5 scenario. A study by Gao et al. [41] assessed the performance of ERA-Interim reanalysis-driven RegCM4 simulations over the period of 1990-2010. Their results showed that this model was capable of reproducing present day climatology over China. The results of June to August were performed better than December to February, when compared to observations. A comparison of biases of mean climatology and extremes �nd similarities in the reanalysis-driving model and the RCM outputs. This indicates that the performance of the RCM to produce mean seasonal climate and extreme daily rainfall is dependent on the performance of the driving models. The dependence and dynamics of the relationship between the boundary condition model and the implementation of the 29 Chapter 3 Case Study on Wanzhou County, China dynamic downscaling through an RCM is depicted in the estimation of mean monthly precipitation over the YRB, shown in Figure 3.11. Figure 3.11: Mean monthly precipitation (mm) over the Yangtze River Basin from 1990 to 2010 derived from observations (CN05 in black), reanalysis model (ERA-Interim in blue) and the RCM model (RegCM in red). Cited from Gao et al. [41]. The performance of the RegCM in estimating seasonal precipitation over the YRB shows that it follows the trend of overestimation by the reanalysis model when compared to the CN05 observations. This overestimation is evident between May to July, while coming closer to the ERA-Interim results from July to August. Additionally, the spatial correlation of precipitation of the model simulations, compared to the observations indicated an insigni�cant relationship. The limitation of RCM results based on the driving model is illustrated in these results. This underlines the importance of an ensemble approach highlighted by previous studies on assessing climate impact on shallow landslide susceptibility [3,8]. The inclusion of results variation of driving models as input to RCM is an important consideration in this case study. Recognizing the importance of selecting an driving model for dynamic downscaling with RCMs, the study on hydrological projections for the Yellow River and Yangtze River in China determined the Hadley Global Environment Model2-Earth System (HadGEM2ES) gave adequate representation of present-day climate based on the validation in previous studies over China [42]. The �ndings from the projections for 2040-2060 for three RCP scenarios and utilizes the China Meteorological Forcing Dataset for 30 Climate Impact Assessments on the Upper Yangtze River Basin Section 3.4 observations for 1980-2000. Though the study focuses on the Tibetan Plateau, in western region of China in Figure 3.8, the performance of this GCM-RCM combination was determined to adequately represent the dataset. This presents an alternative driving model to those utilized in Huang et al. [38], and suggests the options to expand e�ectiveness of an ensemble approach to assessing future climate impact. The research on climate change models over China and in the YRB regions, have to this point highlighted the advantage of deriving insight from high resolution RCMs and the dependence and limitation of the results based on the boundary conditions from the driving GCM models. An alternative model that has been used to assess climate impact across China is the Regional Climate Model (REMO), developed at the Max-Planck-Institute for Meteorology, and maintained by the Climate Service Center Germany (GERICS). Jingwei et al. [43] assess the performance of the REMO over China from 1989-2008 using the ERA-Interim reanalysis. Their �ndings indicate that the REMO model is accurately able to re�ect the distribution of total precipitation, and closely represents the temperature in the summer and total precipitation in the winter. Although �ndings showed a low spatial correlation of results for precipitation, while maintaining an annual mean bias of 300m across China. It was determined that the REMO performance is challenged by complex terrains across China. A further study conducted by Xu et al. [44] utilized a similar combination of ERAInterim and REMO simulation results from 1980 to 2012 to assess the impact of orography in the Hengduan Mountains on downstream the climate of East China. This study focuses on wind circulation but provides valuable insight on the performance of the REMO model on dynamics that a�ect precipitation. The REMO simulations perform better in the summer when the westerly winds from southern China are less than the southwesterly prevailing winds in the winter. It was determined that the simulation overestimated the �ow over the Hengduan Mountains and increased the transport of water vapor downstream. Therefore, these dynamics in circulation are responsible for excessive precipitation, a decrease in cloud cover and an overestimation of temperature. The downstream e�ects of the dynamics of the Hengduan Mountain in the REMO simulation a�ect the domain of East China and thus could in�uence the climate simulations and projections for Wanzhou County. The presentation of the performance of various GCM and RCM combinations are discussed in this section to establish a perspective of performance and limitations in assembling an ensemble for this research on Wanzhou County. The ensemble approach to climate projections was implemented by Gu et al. [45] to analyze the changes in hydrological extremes for projections of 2020-2049 under the RCP 4.5 and 8.5 scenarios. 31 Chapter 3 Case Study on Wanzhou County, China The study utilized an ensemble of 5 RCMs from the CMIP5 outputs, but were only driven by the HadGEM2-AO GCM. The historical annual precipitation was best simulated by the HadGEM2-AO GCM, based the comparison of Gu et al. [46] on the performance of 26 CMIP5 GCMs. The resulting annual precipitation of 2020-2049 increases by 7.65%, compared to the period of 1980-2005 under the RCP 8.5 scenario. The multi-model ensemble projected increases in extreme stream �ow due to intense short-period precipitation events. The 5-member RCM ensemble projected slightly higher values, compared to the results of a similar 27-member GCM ensemble. This study illustrates the implementation for an RCM ensemble for hydrological applications to assess climate change impacts in the YRB. The discussion of related climate literature over the study area suggests that the multi-model ensemble mean in this study could have been heavily in�uenced by the selection of one driving GCM. The consideration of more than one GCM in constructing a multi-model ensemble could thus avoid a potential overestimation in the ensemble mean projection. 32 4Methodology The methodology chapter presents the details of the conceptual framework applied to investigate landslide triggering rainfall conditions and derive climate change projections from regional climate models. The methodology established in this pilot case study area connects extreme statistics techniques from hydrology with atmospheric model correction methods to derive projections of future scenarios, and quantify the uncertainty in these projections. 4.1 Methodological Framework The methodological framework presented in Figure 4.1 illustrates the three phases involved in this study to create shallow landslide-triggering rainfall scenarios. The �rst phase of this study incorporated the landslide inventory available for 19952005 to reconstruct the antecedent rainfall and event rainfall based on the China Meteorological Forcing Dataset (CMFD) precipitation data set [47]. A present rainfall analysis was then carried out to explore the extreme daily rainfall using the CMFD data set. This analysis involved reconstruction of the seasonal antecedent rainfall conditions for the summer season (June-August) by accounting for average rainfall from May-July. The projected rainfall analysis utilized model outputs from the Regional Climate Models (RCM) from the Coupled Model Intercomparison Project 5 (CMIP5) outputs [48]. The process to derive projected rainfall involved bias correction of RCM outputs for daily rainfall and extreme daily rainfall. Finally, climate change factors were computed to apply the present rainfall scenarios and produce quanti�ed future triggering rainfall scenarios. The methodology was implemented across the Wanzhou County area, and procedure is discussed in this chapter. 4.2 Precipitation Data and Climate Model Outputs 4.2.1 Gridded Precipitation Observations In order to understand the interaction between climate change and the magnitude of shallow landslide-triggering rainfall conditions, a reconstruction of rainfall using the dates of the landslides, and the coordinates from the inventory was conducted. Due to the spatial variation of the inventory events, the analysis of rainfall within 33 Chapter 4 Methodology Figure 4.1: Conceptual methodological framework of this research project. the study area utilised gridded precipitation data to capture the spatial characteristics and distribution of rainfall conditions. This section will discuss available rainfall data products, and the approach to modeling the frequency distribution across the study area. A precipitation dataset representing the spatial variability of the study area was determined to be essential, given the regional scale of the area considered, and dispersed locations of the occurrences in the landslide inventory. The sole rain gauge available for this study would not be su�cient to reconstruct the rainfall conditions during and prior to each shallow landslide occurrence. Combinations of gridded precipitation datasets derived from satellite observations and interpolated rain gauge data were taken into consideration to analyse the rainfall-triggering characteristics within the study area. Several satellite-derived precipitation estimate products were evaluated over the Yangtze River basin by Zhe et al. [49]. The conclusion of this evaluation was the ability of the TRMM Multi-satellite Precipitation Analysis (TMPA) research product (3B42V7) to perform adequately in estimating high rainfall intensity event rainfall. The more recent Global Precipitation Measurement (GPM) mission’s Integrated Multi-satellite Retrievals of GPM data (GPM IMERG) over the Yangtze River Basin was determined to have better ability to estimate precipitation at a monthly time scale compared to the daily time scales [42]. Although the TMPA satellite product has considerable performance to assess the spatial characteristics of rainfall in the study area, the availability of TMPA estimates, beginning in the year 1998, was not compatible with the inventory of landslides that began in 1995 [50]. Two rain-gauge derived gridded precipitation datasets were assessed to obtain 34 Precipitation Data and Climate Model Outputs Section 4.2 spatially-distributed rainfall estimates within the study area. The �rst dataset is the daily gridded precipitation dataset at 0.25 °× 0.25 ° resolution under the Asian Precipitation— Highly Resolved Observational Data Integration Towards the Evaluation of Water Resources (APHRODITE) project. The APHRODITE gridded precipitation dataset was estimated through the interpolation of a dense network of rain gauge observations from 1961-2005 across Asia [51]. An updated APHRODITE-2 improves the ability of the dataset in analysing extreme precipitation events through the integration of multi-satellite merged precipitation products and the ERA-Interim reanalysis to ensure uniformity in the daily accumulation of precipitation across domains [52]. The assessment of Zhenyu & Tianjun [53] on the APHRODITE product when compared to rain gauge station data shows accurate seasonal precipitation characterization in mean states. However, it was concluded the APHRODITE dataset underestimated precipitation intensity and showed signi�cant di�erence in spatial patterns of intensity trends, when compared to station data. The second precipitation dataset, which was selected for study, was from the China Meteorological Forcing Dataset (CMFD) with a temporal resolution of 3 hours and a spatial resolution of 0.1 ° . This data set was derived from the integration of multi-satellite remote sensing products, reanalysis datasets and rain gauge observations from 1979 to 2018. The algorithm for the creation of the precipitation dataset utilised a monthly scale for interpolation, due to a smoother spatial distribution compared to results from sub-daily interpolation. The sub-daily values were obtained through the integration of Chinese Meteorological Agency (CMA) station observations to gridded data from the Global Land Data Assimilation System (GLDAS) and Tropical Rainfall Measuring Mission (TRMM) at 0.1 ° . The interpolated sub-daily precipitation rate was then proportionally adjusted to the interpolated precipitation estimates at a monthly scale to obtain the spatial and temporal characteristics described above [47]. Daily precipitation at a spatial resolution of 0.1 ° was selected from the CMFD dataset for the rainfall analysis and reconstruction of inventory precipitation characteristics for this study. The spatial variability of the shallow landslide inventory depicted in Figure 4.2, together with centers of the CMFD pixels highlight the importance of the gridded precipitation dataset in contrast to the sole rain gauge made available for this study area. The spatial coverage and resolution of the CMFD data set was determined to give better representation of precipitation conditions of the nearest corresponding inventoried landslide. 35 Chapter 4 Methodology Figure 4.2: Spatial distribution of shallow landslides, center points of the CMFD gridded observations, and the location of the available rain gauge. 4.2.2 Multi-model Climate Projection Ensemble In order to determine the change in magnitude and precipitation of landslide triggering, as proposed in the research questions, this study utilized projections from climate change models part of the outputs from Phase 5 Coupled Intercomparison Modeling Project (CMIP5) outputs [54], under the Coordinated Downscaling Experiment (CORDEX) Project. These climate model outputs enabled the development of rainfall scenarios derived from the projected changes in precipitation. The climate model outputs used in this study consist of model outputs from Global Circulation Models (GCMs) that simulate emission scenarios to represent changing boundary conditions de�ned by Representative concentration pathways (RCPs). Di�erent RCPs represent di�erent cases of optimistic (RCP 2.5), moderate (RCP 4.5), and extreme (RCP 8.5) scenarios in emission releases with the objective of understanding possible outcomes based on insight from earth system dynamics modeled in GCMs [55]. A particular scenario of interest in this research project is the RCP 8.5 scenario. This 36 Precipitation Data and Climate Model Outputs Section 4.2 is the scenario that assumes high population growth, low income and slow rates of technological adaptation that will lead to long-term increases harmful emission in the absence of climate change policies [56]. This scenario the climate interaction with socioeconomic driving factors that could result conditions that will increase shallow landslide risk. One limitation of large-scale simulations covering long time scales is the spatial resolution of information on the projections. While GCMs simulate climate dynamics to the spatial order of magnitude in the range of 100 kilometers, the projection information is likely incompatible with site-speci�c assessments that require higher spatial resolution to analyze changes in precipitation regime. The introduction of dynamic downscaling with Regional Climate Models (RCMs) aimed to provide higher resolution projection information through taking the climate output from the GCMs as boundary conditions to locally simulate the atmospheric physics at �ner spatial resolutions between 10-20 km. A combination of GCMs and RCMs were considered in this methodology to develop a multi-model ensemble projection. An overview of the climate studies conducted in Section 3.2, and the review of the state of the art in climate change assessments for shallow landslide susceptibility, discussed in Section 2.3, suggested the ensemble approach provides more robust and comprehensive climate change projections. Giorgi et al. [57]�nd that the utilization of a multi-model ensemble is a necessary approach to characterize uncertainties in climate projections. Thus, this research project considered an ensemble of 4 members considering 2 GCM and 2 RCM outputs to pilot the methodological framework, presented in Figure 4.1. This research project utilizes a relatively small ensemble, where Giorgi & Coppola [58] recommend that a minimum of 4 to 5 RCM models are necessary to obtain robust precipitation projections. Further research on the optimal multi-model ensembles of GCM on the other hand suggests that 9 GCMs can adequately represent the climatology over China [52]. In assessing the importance of ensemble members, a study on exploring performancebased weighting for RCMs by Christensen et al. [59] found no compelling evidence for improved projections of this method when compared to the application of equal weights. Advances in building robust projections through the reliable ensemble averaging (REA) method has demonstrated ability to reduce uncertainty and address systematic errors in RCMs to produce more reliable hydrological impact assessments [60]. In order to accomplish the research objectives in this project, a simpler equal weight approach was adopted, but the application of the REA method to improve the ensemble projections was taken into consideration for future research recommendations. This methodology was proposed to answer the research question, 4 ensemble members were 37 Chapter 4 Methodology The CVM test together with the AD test are estimated in practice to measure the goodness-of-�t with a weighted emphasis on the tails of the distribution Laio [69]. The CVM tests were implemented by the ‘goftest’ R package. 4.5 Climate Change Analysis This section presents the methodology developed to perform bias correction on the climate change models, assess the correction’s performance through cross-validation, and derive a climate change factor based on projections for extreme daily rainfall and mean seasonal rainfall independently. Two climate change factors are derived from the bias-corrected data to represent the climate signal from the projections applicable to extreme, and seasonal rainfall. The climate change analysis aimed to produce multi-model ensemble projections under RCP 8.5 conditions for mean seasonal rainfall and extreme daily rainfall. The scenarios are de�ned by the future periods in the Mid 21st Century from 2021-2060 and Late 21st Century from 2061 2100. The mean ensemble projections were reported as Climate Change Factors (CCF) relative to the reference period of 1979 to 2018. 4.5.1 Bias Correction Methodology This section details the procedure implemented to separate the climate model outputs before bias-correction for extreme daily rainfall, to represent triggering event rainfall, and mean seasonal rainfall, that will represent antecedent rainfall. The di�erent bias corrections through empirical and parametric quantile delta mapping transfer functions were implemented. This resulted in a climate change factor applied to obtain the mean ensemble projections that represented extreme daily rainfall and mean seasonal rainfall. An overview of the framework for the bias correction methodology is presented in Figure 4.4. The bias correction proposed utilized was the Quantile Delta Mapping (QDM) method. Two separate QDM procedures were shown in Figure 4.4 to correct the outputs and extract the climate signals for extreme events and antecedent events as independent variables. The bias correction framework enabled the correction of systematic errors in the climate model outputs. The results yielded bias-corrected results for the Mid-21st Century and Late 21st Century scenarios, corrected against the reference period of 1979 to 2018. The dataset of the reference period consisted of simulations from the historical model results between 1980 to 2005, and the RCP 8.5 projections from 2006 to 2018. The de�nitions of the climate change scenarios are summarized in Table 4.2. 44 Climate Change Analysis Section 4.5 Figure 4.4: Flowchart of data required and bias correction methods and work�ows to derive climate change projections for extreme daily rainfall and antecedent rainfall. Table 4.2: De�nitions of climate change scenarios. Scenario Period (Years) RCM Experiment Outputs Reference 1979-2018 Historical + RCP 8.5 Mid 21st Century 2021-2060 RCP 8.5 Late 21st Century 2061-2100 RCM 8.5 The reference period consisted of simulations from the historical model results between 1980-2005, and the RCP 8.5 projections from 2006 to 2018. The underlying assumption in the decision to combine the projection and historical model results to establish a combined reference period dataset was that the RCP 8.5 scenario is representative of the conditions under which the CMFD observations were recorded. This may not necessarily be the case, but the essence of this methodological framework is to capture the trend in the bias-corrected precipitation in between time periods from the reference to future climate scenarios. Thus, supplementing the historical simulation results with future projection results to establish a reference distribution for bias correction will account for potential systematic bias and uncertainty in early periods of the RCP 8.5 projections. 4.5.2 �antile Delta Mapping Bias correction is performed on climate modes to address the systematic distributional biases in precipitation outputs. The assessment of present day climate by Gao et al. [40] indicated that high resolution RCM simulations critically outperform GCM results in reproducing the climate over China, with a speci�c advantage in assessing the development of the East Asian monsoon. 45 Chapter 4 Methodology Research on the necessity and advancements of di�erent bias correction methods on impact assessments and climate change projections for the YRB and over the entire China are widely studied in contemporary literature [37–39,71]. A common approach to perform this task is through Quantile Mapping (QM). This method for bias correction utilizes a transfer function derived from precipitation observations to correct the frequency distribution in the climate model’s historical outputs. The transfer function can be derived empirical from the CDF of the observations or with parameters of a �tted a frequency distribution model. The bias correction of a variable through QM is given by the Equation: ˆ G=1 >,2[<,?(G<,?)] (4.13) where G12 is the bias corrected precipitation, 1 >,2 is the inverse CDF derived from the observations, <,? is the CDF derived from the model projections, and -<,? is the projected precipitation to be corrected. The limitation of the QM method is that the future model projections are corrected using the information derived from the historical period. Trends and information about the future model outputs are not considered and lost in the bias correction procedure. In order to address this limitation Cannon et al. [72] proposed the use of a Quantile Delta Mapping (QDM) bias correction method to perform bias correction while preserving the changes in quantiles in the projected climate outputs. Tong et al. [37] described the methodology for QDM that was performed on RegCM4 model projections over China. The procedure for QDM follows the process of �rst detrending the simulated model projection through QM during a calibration period, given by Equation (4.13). The relative changes in quantiles between the calibration period and the projected period of interest was then derived from the ratio between the distribution of precipitation. This ratio of change is derived from the Equation below: J(C)= G(<,?) 1 <,2[(C) <,?(G<,?(C)] (4.14) where J(C) is the factor of relevant change in precipitation for the period of interest C , 1 <,2(C) is the inverse CDF derived from the model projections over the reference period, (C) <,? is the CDF derived from the model projections over the period of interest C , and (G<,?(C)is the precipitation model projections over the period of interest C. The bias corrected precipitation is derived from the product between the factor of relative change and the detrended variable ˆ Gas shown in the Equation: G12 (C)=ˆ GJ(C)(4.15) where G12 (C)is the bias corrected precipitation time series for the period of interest C. 46 Climate Change Analysis Section 4.5 The objective of performing bias correction on RCM outputs was to correct the substantial systematic error, partly inherited from the GCM on the boundary conditions, and obtain results to assess trends and climate signals. Speci�c applications of climate change studies require insight and analysis from speci�c climate signals. Bias correction was performed separately for daily precipitation and extreme daily rainfall. Empirical �antile Delta Mapping for Daily Rainfall Assessments of trends in the mean climatology will require climate signals covering seasonal and annual measurements. If not corrected separately, the recti�cation of systematic error that might distort the distribution of daily, monthly or annual projections. The QDM bias correction method presented in this section adequately enhanced the insight from the climate models to attain more certainty in the projections. Previous studies have established that a non-parametric transfer function to perform quantile mapping bias correction has the advantage of not requiring an assumption of the distribution of the observed and modeled rainfall [37]. This was deemed appropriate to obtain results of the in�uence of climate change on antecedent rainfall found in the climate projection models. The application of a transfer function derived from the empirical cumulative distribution function would directly account for the distribution of daily rainfall without adding uncertainty derived from a �tted distribution function. The bias-corrected seasonal precipitation would result in a more realistic representation of the distribution observed, and retain the relative change in precipitation from the model, without distorting the output frequency distributions. Therefore, the bias correction of daily precipitation was conducted for the months of May to August to assess the summer season projections and meet the research project objectives. Parametric �antile Delta Mapping for Extreme Daily Rainfall The bias-correction of results to capture precipitation extremes required the application of QDM to involve considerations of the limitation of the climate models to simulate the frequency and magnitude of such events on regional and local scales [73]. The necessity to conduct a bias correction on the simulated climate date using annual maxima instead of the conventional daily precipitation was established in a study by Kim et al. [74]. Their study found that the conventional bias-correction method using daily data was limited in ability to capture extreme rainfall quantiles, and required a bias correction for the application of extreme rainfall frequency analysis. Though RCMs were found to over predict the probability of extreme daily precipitation, it was found that results could be corrected to adequately represent the upper-tail distributions in observations Durman et al. [75]. While this provided promising reassur47 Chapter 4 Methodology ance of the present conditions, it was also found that RCM models have been found to be unable to represent the relationship between extreme rainfall and temperature [76]. Research indicates that the distribution and magnitude of extreme rainfall in the future projections would be subject to more uncertainty, and be less likely to represent reality. Thus, given the limitations of GCM and RCM models to capture extreme precipitation events, it cannot be assumed that an empirically-derived extreme precipitation distribution will adequately represent reality. The alternative to assess the trend in extreme daily rainfall was to capture the trend in extreme precipitation through a parametric QDM approach. Kim et al. [77]�tted General Extreme Value (GEV) distribution models to the annual maxima datasets to perform the QDM. The selection of GEV distribution models was based on the 2019 �ood estimation recommendations of the South Korean Ministry of Environment. The demonstration of the application of parametric QDM to bias correct extreme daily rainfall inspired the direction for bias correcting results of the ensemble members in this research project. The approach to establishing the transfer function for bias correction for extreme daily rainfall is outlined in the methodology for the extreme rainfall analysis in the present, shown in Section 4.4. The Gumbel �t was selected for this area due to the suggested performance in capturing extreme events by Xiao et al. [21]. Similarly, the selection of the GEV distribution by Kim et al. [26] was based on the suggestion �ood estimation guidelines for South Korea, issued by the Ministry of Environment in 2019. The extraction of monthly maxima in the simulation outputs for June to August for the scenarios de�ned by the years listed in Table 4.2 was considered in the application of the QDM bias correction. The magnitude of extreme daily rainfall were then determined for di�erent return periods with Equation (4.5), using the bias-corrected Gumbel �t parameters. The main limitation of the proposed QDM bias correction for extreme daily rainfall is the assumption that a Gumbel frequency distribution will remain adequate under future climate conditions. Huang et al. [38]�nds that the application of a mixture of distribution to perform as a transfer function for bias correction outperforms a single distribution by allowing for spatial variation of rainfall distributions in di�erent grid cells. Although the implementation of the parametric QDM resulted in biased-corrected estimations of Gumbel frequency distributions to represent the extreme daily rainfall in the future scenarios. This potential improvement by applying a mixture of extreme value distributions was considered to bear potential in improving the results of this initial pilot study. 48 Climate Change Analysis Section 4.5 4.5.3 Cross-Validation for Bias-corrected Daily Rainfall The objective of the daily rainfall bias correction was to enable the assessment of antecedent rainfall through the derivation of mean seasonal rainfall. Thus, it was important to assess the performance of the bias correction results in determining the temporal accuracy and magnitude of error in daily precipitation simulations. In order to assess the performance of the bias corrected climate simulations, a 3fold cross-validation scheme was implemented. The evaluation of bias correction methods on models through cross-validation was implemented in the previous studies with calculation of the Mean Absolute Error (MAE) and the Root-Mean-Squared Error (RMSE) to measure performance [77,78]. This study adopted a cross-validation scheme particular to the bias correction approach for daily rainfall, as shown in Figure 4.5 Figure 4.5: Flowchart of cross-validation procedure from splitting the data into training and testing to deriving performance metrics. 49 Chapter 4 Methodology The approach was implemented by randomly sub-dividing CMFD observations and RCM outputs by 2 / 3for training and 1 / 3for testing. The empirical QDM bias correction method for daily rainfall was then performed. The training datasets from the CMFD observations and the RCM outputs would be references to establish the transfer functions, and correct the sub-divided testing dataset from the RCM output. The performance was assessed through testing the accuracy and magnitude of the bias-corrected RCM testing dataset versus the CMFD observation testing dataset. The performance was measured by MAE and RMSE, respectively. This procedure was performed for random subdivisions of data sets over 100 iterations to adequately assess the performance of the empirical QDM method in correcting daily rainfall data. The strength of the k-fold cross-validation procedure is in its robustness to assess the uncertainty and deviation of climate models outputs from observations. Though critics of this method cite that long-term trends and oscillations in dynamic climate systems may lead to random realization of long-term climate variability. This would render the results of cross-validation versus observations a result of internal variability, and be unable to provide sensible insight of the bias correction [79]. The downstream dynamics of the Hengduan Mountains, and the in�uence from interaction of the Indian and East Asian Monsoons are internal factors of variability present that a�ect Wanzhou. The insight derived from the cross-validation procedure, presented in Figure 4.5, was a measure of the uncertainty and variability of the biascorrected results. An expanded analysis incorporating spatial aspects and statistical signi�cance of the bias-corrected results could lead to more certain measures of performance for the QDM bias correction method. 4.5.4 Validation for Bias-corrected Extreme Daily Rainfall The objective of the parametric QDM method for extreme rainfall was to perform a model bias correction through the comparison �tted Gumbel frequency distribution models. The evolution in magnitude and the distribution of the Gumbel �tted representation extreme daily rainfall is the intended outcome of this bias correction methodology. The validation procedure for the extreme daily rainfall was performed through the resulting ordered maximum monthly precipitation from the bias-corrected dataset with the counterpart dataset from the CMFD observations for the months of July to August in the reference period of 1979 to 2018. The performance for the bias correction of extreme daily rainfall was measured by Mean Absolute Error (MAE) and the Root-mean-squared Error (RMSE) of the ordered values. The intention of these error measurements was to capture the accuracy and magnitude of the bias-corrected results. Additionally, a Pearson Correlation Coe�cient 50 Climate Change Analysis Section 4.5 was derived to assess the ability of the RCM simulation results to capture the extreme daily rainfall in the CMFD observations. Finally, a KS test was performed to assess the ability of the bias correction method to correct the simulation results. This was performed to test if frequency distribution of extreme daily rainfall belonged to the same distribution as the observations within a con�dence interval of 95%. The KS test performed followed the description in Section 4.4. 4.5.5 Climate Change Factors In order to incorporate the trend of climate signals for extreme daily rainfall and mean seasonal rainfall, a Climate Change Factor was derived, following the relationship given the Equation (4.16). ⇠⇠8= %DCDA4,">34; %8BC>A820;,">34; (4.16) where ⇠⇠8 is the the e�ective climate change factor of or either the value of the mean seasonal rainfall, ⇠⇠"(' , or extreme daily rainfall, ⇠⇠⇢⇡',) , for a return period, ). The derivation of the ⇠⇠⇢⇡',) for extreme daily rainfall considers, %DCDA4,">34; , as the magnitude of rainfall on the return period curve derived from the future scenario simulations corresponding to a return period ) , and %8BC>A820;,">34; , as the magnitude of rainfall on the return period curve derived from the historical simulations corresponding to return period return period ). The derivation of the climate change factor, ⇠⇠"(' , for antecedent seasonal precipitation is computed through considering the mean seasonal antecedent precipitation, taking %DCDA4,">34; as the mean seasonal rainfall from future scenario simulations and %8BC>A820;,">34; as the mean seasonal rainfall derived from historical scenario simulation. A ⇠⇠"(' value was was derived for future scenario simulations de�ned by Table 4.2. 51 5Results and Discussion This chapter presents the results and discusses the �ndings of this research project. The results are divided into three parts, following the presentation of the methodological framework in Figure 4.1. The �rst main section in the results discuss the �ndings from the reconstruction of historical triggering rainfall conditions. An analysis of the present mean seasonal rainfall patterns and the Gumbel distribution model of extreme daily rainfall then discussed. Finally, results of the bias correction of RCM outputs, and the climate change analysis derived from the ensemble projections are assessed. 5.1 Part 1: The Reconstruction of Triggering Rainfall Conditions This �rst part of this research project was the reconstruction of event rainfall and antecedent rainfall conditions corresponding to the shallow landslides recorded in the inventory for the period of 1995-2005. The rainfall was extracted from the CMFD gridded precipitation observation and recharge estimates were applied to attain the e�ective antecedent recharge conditions. An analysis of the reconstructed rainfall conditions during the summer seasons of 1995 to 2005 and an assessment of their reliability was then carried out. Temporal uncertainty in the recorded landslide date was detected for landslides occurring during periods with no estimated e�ective antecedent recharge that showed a magnitude of zero event rainfall. The analysis of the sensitivity of the inventory dates to a time-shifting algorithm implemented to search for maximum triggering event rainfall within a period prior was conducted to address the lack of any rainfall in the reconstruction based on the original recording dates. Through the exploration of uncertainty in the landslide recording dates, it was determined that signi�cant uncertainty until 7 days prior to the original record dates would be a hindrance to the reconstruction of triggering rainfall conditions. Furthermore, an analysis of the sensitivity of antecedent rainfall conditions to this temporal uncertainty was conducted. 53 Chapter 5 Results and Discussion Figure 5.6: Histogram (A) and density distribution (B) of the reconstructed 30-day mean e�ective antecedent recharge for landslides occurring between June to August 1995 to 2005. 60 Part 1: The Reconstruction of Triggering Rainfall Conditions Section 5.1 Figure 5.7: A 2D density distribution of Event Rainfall versus the 30-day E�ective Antecedent Recharge from June to August 1995 to 2005. Figure 5.7 depicts the 2D density distribution of the reconstructed e�ective antecedent recharge and event rainfall. The concerns over uncertainties in the data and the ability to reconstruct triggering rainfall conditions are highlighted in the cluster of landslide occurrence seen with a range of e�ective antecedent recharge of 0 mm/day to 1 mm/day, and event rainfall between 0 mm/day to 10 mm/day. This cluster of occurrences highlights the uncertainties in the data sets used to reconstruct the shallow landslide triggering rainfall conditions. The observation of this cluster prompted a deeper analysis in the potential temporal uncertainty in record date as a feasible initial correction step to be able to utilize the inventory and proceed with reconstructing the hydro-meteorological conditions for shallow landslide triggering in the study area. Although a large cluster of landslides Figure 5.7 occur with no apparent hydrometeorological forcing acting as the triggering mechanism, a cluster of landslides occurs with a range of e�ective antecedent recharge of 0 mm/day to 1 mm/day and event rainfall between 15 mm/day to 25 mm/day. This cluster of landslides represent reasonably reconstructed triggering rainfall conditions that can be utilized in the calibration of models for a susceptibility analysis. 61 Chapter 5 Results and Discussion Figure 5.8: A comparison of the e�ect of the time-shifting algorithm (in red circles) for 2 to 11 day periods (left to right, top to bottom) and the di�erence between the original record dates (in blue triangles) with focus on reconstructed rainfall conditions occurring during summer season months from 1995-2005 with no estimated recharge. 62 Part 1: The Reconstruction of Triggering Rainfall Conditions Section 5.1 5.1.2 An Assessment of Temporal Uncertainty in the Inventory The potential for uncertainties to exist in shallow landslide inventories is common and problematic for the assessment of shallow landslide susceptibility. This is especially problematic since landslide inventories and known landslide occurrences are utilized to calibrate models used to generate triggering thresholds and spatial predictions. The resulting clusters of landslides with no signi�cant triggering hydro-meteorological conditions in the initial reconstruction of triggering rainfall conditions, it was determined that the most feasible solution for this study to address the general uncertainty in the inventory would be to assess the temporal uncertainty in the record dates. Other uncertainties in the data sets used to reconstruct the triggering rainfall conditions such as the spatial uncertainty of the locations of the landslides and the uncertainty in the ability for the CMFD gridded precipitation reanalysis to capture intense localscale rainfall after the spatial interpolation and processing were identi�ed as limitations that require further time and resources to address. As identi�ed in Figure 5.7, the cluster of landslides showing no e�ective antecedent recharge and no event rainfall was selected and put in focus for the assessment of temporal uncertainty in the inventory. The approach assumed that due to operational constraints, and procedure for recording sighted shallow landslides by the personnel recording the occurrences, it is possible that the record dates stated in the inventory were the dates in which these personnel sighted the landslide, and not the actual dates of failure. Therefore, this temporal uncertainty in recording dates was assessed using a time-shifting approach that searched for the maximum event rainfall in the CMFD precipitation observations of the pixel representing the rainfall conditions of the landslide location within a period of 1 to 11 days prior to the record date found in the inventory. The implementation of this time-shifting solution to address temporal uncertainty was assessed through the examination of critical 0 mm/day e�ective antecedent recharge that represented 37 landslides, or 25% of the reconstructed inventory. The results of the time-shift algorithm’s search for a maximum event rainfall compared to the original inventory date is shown Figure 5.8. The shift in the magnitude of event rainfall observed after a 2-Day and 5-Day shifting window was applied. This revealed shifts in position from 0 mm/day of event rainfall towards 20 mm/day of event rainfall for 10 occurrences. There was no shift in record date of landslides around the 20 mm/day threshold observed. The change in position of dates resulting in event rainfall placed in the 20 mm/day region was also observed in a 7-day period. These results suggest that the presence of the temporal uncertainty in the originally reconstructed rainfall data set that was based on the initial recording date in the inventory. Although an agglomeration of event rainfall shifted to a range within the 20 63 Chapter 5 Results and Discussion mm/day threshold is observed consistently from a 2-Day to 7-Day shift, the expansion of the time shifting period to 11 days reveals expanded uncertainty with larger magnitudes of shifts surfacing. This indicates the possibility of an even longer period of uncertainty and more signi�cant and higher magnitudes of event rainfall during months estimated to have zero e�ective antecedent recharge. Figure 5.9: A comparison of the percentage of the inventory a�ected by the detection of a maximum event rainfall for di�erent periods of days prior to the record date The assessment of the time-shifting solution was expanded from its focus on landslide occurrences with no e�ective antecedent recharge to all the occurrences in the inventory. The percentage of data points shifted from their original record dates based on maximum event rainfall found in the CMFD data set is presented in Figure 5.9 for the time-shifting windows of 1 to 11 days prior to the record date. The results that a shift to �nd a maximum event rainfall of one day changes 88% of the event rainfall based on the original recording dates. There is a noticeable increase in the number of landslides a�ected in the application of a maximum event rainfall search algorithm in the data set from a 1-Day search period to a 7-Day period, increasing from 88% to 96% of the data set changing. Between 7 and 11 days, there is a di�erence of 2%, from 96% of the data set changing after a 7-day search period is applied to 94% with an 11-day period. It can be observed that a period of relative stability between a 4-Day and a 5-Day 64 Part 1: The Reconstruction of Triggering Rainfall Conditions Section 5.1 search period suggests that no signi�cant change in data points when searching for the maximum triggering event rainfall. The results of this assessment of the sensitivity of the reconstructed event rainfall indicates large uncertainty in the record date and its representation of the day the recorded shallow landslides were triggered. Based on the results of this assessment, the implementation of a 7-Day window to represent maximum uncertainty in the event rainfall and best represent the reconstruction of landslide triggering hydro-meteorological conditions was further assessed. Figure 5.10: An inter-comparison between the percentage of the inventory a�ected by the detection of maximum event rainfall between paired intervals of consecutive days prior to the record date. In order to support the adequacy of a 7-Day period prior to the original record date in the inventory to reasonably represent the uncertainty at the maximum and adequately allow insight from the reconstruction of rainfall conditions, the di�erence between time periods of shifted data sets was assessed. This was intended to examine the number of landslide occurrences a�ected between time-shifting periods to understand 65 Chapter 5 Results and Discussion the movement of data points towards a maximum event rainfall and the stability of each period of search. The intercomparison of inventory data points a�ected between each time-shifting period previously analyzed is presented in Figure 5.10. In contrast to Figure 5.9, this analysis of the movement of data points and the search for a probably triggering event rainfall reveals a point of relative stability between the time-shifted periods of 2-3 days and 3-4 days. In these 3 periods, 20% of data points changed when an algorithm to search for the maximum event rainfall was applied. When compared to the original record dates, this period sees 91% to 94% of data points increase. In contrast to the observed stability of the 4 to 5 day search period in the Figure 5.9, the intercomparison of data points in Figure 5.10 indicates that between the 4-5 day search period reveals that 63% of the data points were shifted. This relative stability is not consistent in Figure 5.9 and Figure 5.10. The 7-Day search window emerged as the period of maximum uncertainty, where a comparison to the 5-Day window yielded 91% of data points changing in the search for a maximum event rainfall. While the change between a 7 and 11-day search period was less prominent, with 59% of data points �nding a new maximum event rainfall, the uncertainty is still signi�cant. The �rst assessments implemented in this section aimed to understand the sensitivity of the rainfall conditions reconstructed around the original record dates to a timeshifting solution to search for the maximum daily rainfall. The second assessment established the relative stability of the data sets created by the maximum daily rainfall search algorithm for di�erent periods of uncertainty, in days prior to the original record date. These assessments were utilized to assess the temporal uncertainty and the adequacy of a time-shifting solution to address the temporal uncertainty in the inventory. The sensitivity analysis presented indicates that the 7-Day period represented conditions of maximum uncertainty. Therefore, this maximum period of uncertainty was taken into account and used to further analyze the reconstructed triggering rainfall conditions that this part of the research project aimed to achieve. Incorporating a 7-Day Period of Uncertainty The resulting assessment of temporal uncertainty in the inventory resulted in the selection of a 7-day period to reconstruct the triggering rainfall conditions This period represents the maximum uncertainty when considering a period of 1 to 11 days prior to the originally recorded inventory dates. The critical assessment of uncertainty began with shallow landslide events that occurred during periods with no estimated recharge. The reconstructed data set reveals groups of data that, despite the time-shifting solution, show negligible event rainfall as the triggering mechanism. Consistently 66 Part 1: The Reconstruction of Triggering Rainfall Conditions Section 5.1 observed with the reconstruction around the original record dates is an apparent gravitation of landslides around the 20 mm/day threshold of triggering event rainfall. A group of landslide events occurring from July to August 2002 suggest that operational constraints in identifying and recording landslides manifested in the delay of recording to the inventory. The study area covers 3,457 km and identifying and monitoring shallow landslides in this region is a challenging task. Figure 5.11: A 2D density plot of the event rainfall and the antecedent rainfall with a 7-day uncertainty period applied occurring from June-August 1995-2005. The implementation of the 7-Day time-shift solution to account for the temporal uncertainty in record dates was applied to the entire data set. The 2D density plot of the reconstructed antecedent rainfall and event rainfall magnitudes incorporating the record date uncertainty is presented in the Figure 5.11. Similar to 2D density plot of the reconstructed triggering rainfall conditions, based on the original recording dates in Figure 5.1, the antecedent rainfall estimates in the main cluster of landslides in Figure 5.11. This suggests that the occurrence of shallow landslides correspond to a range of antecedent rainfall from 3 mm/day to 6 mm/day. While the shift in rainfall conditions suggest a concurrent occurrence with the event rainfall of 10 mm/day and 30 mm/day. In contrast to the results of the reconstructed rainfall conditions without the incorporation of record date uncertainty, a cluster of 67 Chapter 5 Results and Discussion event rainfall and antecedent rainfall conditions is more clearly depicted with event rainfall from 60 mm/day to 90 mm/day and antecedent rainfall between 8 mm/day and 11 mm/day. The graphical interpretation of the relationship between antecedent rainfall and event rainfall conducted in this section may add insight to establishing thresholds of occurrence and suggest ranges of rainfall by which shallow landslide susceptibility modelers can begin their risk assessments of the study area. Figure 5.12: A 2D density plot of the event rainfall and the e�ective antecedent recharge resulting 7-Day uncertainty period occurring from June-August 1995-2005. The analysis of the reconstructed rainfall conditions incorporating temporal uncertainty in the record dates was extended to the estimated antecedent rainfall. The 2D density plot of the reconstructed data set is shown in Figure 5.12. The dominant cluster of landslides assessed in Figure 5.11 shifted downward towards the x-axis, re�ecting that a signi�cant number of shallow landslides occurred during months estimated to have no e�ective antecedent recharge. This cluster suggests a majority of landslides occur with e�ective antecedent recharge conditions between 0 mm/day and 1 mm/day and accompanied by a wide spread of event rainfall between 5 mm/day and 30 mm/day. Two clusters of landslide occurrences also formed in the visualisation of density in the 68 Part 1: The Reconstruction of Triggering Rainfall Conditions Section 5.1 data set, indicate potential combinations of rainfall conditions that can be considered as higher magnitude event rainfall. The �rst cluster is located with a range of e�ective antecedent recharge between 0 mm/day and 1 mm/day, and event rainfall between 60 mm/day and 90 mm/day suggests a group of landslides that were triggered by failure of the soil column through the vertical �ow process. The second group of clustered landslides suggests that signi�cant e�ective antecedent recharge more than event rainfall triggered the shallow landslides. This cluster is located with e�ective antecedent recharge of 1 mm/day to 1.5 mm/day and event rainfall of 0 mm/day to 10 mm/day. This group of landslides suggests this combination of antecedent rainfall and a minute amount of event rainfall enabled the triggering mechanism through a raised groundwater head by the process of lateral subsurface �ow paired with vertical �ow through the soil column. Although these results can be insightful in approaches incorporating soil mechanics and physics-based triggering mechanisms, from a standpoint of threshold determination and data-driven insight the uncertainty and the wide range of the clustered landslide occurrences does not give clear insight into the triggering mechanisms and scenarios for shallow landslides in the study area. The results of this section suggest that based on the cluster of event rainfall with large values, an assessment and analysis of extreme daily rainfall distributions across the study area could provide valuable insight into the modeling shallow landslide susceptibility under extreme conditions. 69 Chapter 5 Results and Discussion 5.1.3 An Analysis of the Reconstructed Antecedent Rainfall The results presented throughout this chapter have established that the temporal uncertainty in the record dates of the inventory can be reasonably addressed by incorporating a time-shift solution to incorporate maximum uncertainty within a 7-day window to reconstruct triggering rainfall conditions. An analysis of the in�uence of the incorporation of a 7-day window of uncertainty on antecedent rainfall conditions was also assessed, and the results suggest that a 30-day antecedent rainfall duration can provide a reasonable and robust estimation of antecedent conditions with the incorporation of uncertainty from the inventory. The �nal section of results in this chapter analyzes the distribution of triggering rainfall conditions for summer season shallow landslides based on the precipitation observations derived from the spatially distributed CMFD data set. Figure 5.16: Comparison of the density distributions of antecedent rainfall of the reconstructed inventory with a 7-day window of uncertainty and the distribution antecedent rainfall in CMFD observations from 1995 to 2005. 76 Part 1: The Reconstruction of Triggering Rainfall Conditions Section 5.1 Figure 5.16 compares the distribution of antecedent rainfall of the reconstructed summer season triggering rainfall conditions with the antecedent rainfall observations from all CMFD pixels within the study area. The 30-day average rainfall is derived from CMFD observations across the study area for the months of May-July to represent all possible antecedent rainfall conditions from 1995-2005. The distribution of 30day antecedent rainfall derived from a 7-day time-shift solution is also visualized in Figure 5.16. There is an apparent similarity between the distributions of the two data sets described in the plot of Figure 5.16, with the medians of each distribution, depicted by vertical dashed lines appearing to be apart by a small di�erence. A Wilcoxon Rank Sum Test was performed to assess if the di�erence between the medians of both distributions bore statistically signi�cant similarities. The test statistic comparing both 30-day antecedent rainfall data sets returned a p-value of 0.428, indicating strong statistical signi�cance within a 95% con�dence interval that the location of the medians of both distributions can be considered similar. Furthermore, a comparison of the distribution of the 30-day antecedent rainfall from May-July and May-August identical distributions, and similar medians, shown in Figure 5.16. This indicates that mean seasonal rainfall from May to August can adequately represent the average antecedent rainfall conditions that were likely to have triggered shallow landslides from 1995-2005. The analysis on Figure 5.16 suggests that shallow landslides occurring during the summer season across Wanzhou from 1995 to 2005 were triggered by a range of antecedent rainfall amounts distributed similarly rainfall that fell over the entire study area. The distribution of reconstructed 30-day antecedent rainfall conditions that triggered landslides in the inventory show statistical similarity in mean of that distributed in the observation data of rainfall between May to August. Hence, indicating that in the assessment of shallow landslide susceptibility, the mean seasonal rainfall conditions can adequately represent conditions for shallow landslide triggering, as revealed by the reconstruction of the inventory data set. An important implication of this �nding is that mean seasonal rainfall can be applied as an antecedent condition when modeling shallow landslide susceptibility in the future. 77 Chapter 5 Results and Discussion 5.2 Part 2: An Analysis of the Present Summer Season Rainfall The second part of this research project ful�lls the objective of analyzing the present extreme daily rainfall and the mean seasonal rainfall during the summer in Wanzhou County, China. The objective of the results of this process is the establishment of a reference scenario in the present by which climate change signals from climate projections can be applied to. The period of analysis of this section is between 1979 and 2018, corresponding to the duration by which the CMFD gridded precipitation observations are available for. Using these gridded precipitations, the mean seasonal rainfall from May to August, is analyzed. After which, extreme daily rainfall (EDR) events during this season are extracted using a block maxima approach. Using the extracted EDR within the season, the frequency distribution is modeled through �tting a Gumbel distribution. The extreme frequency distribution �ts are implemented for all CMFD pixels within the study area and validated against three goodness-of-�t tests, namely the KolmogorovSmirnov (KS) Test, the Anderson-Darling (AD) Test, and the Cramer-von-Mises (CVM) Criterion Test. The spatial distribution of the extreme daily rainfall and an analysis of the return period curves within the study were then analyzed. The resulting spatially distributed information on mean seasonal rainfall and extreme daily rainfall distributions were therefore established as the reference rainfall scenarios for the assessment of future climate change projections. 5.2.1 A Spatio-Temporal Analysis of Mean Monthly Rainfall The landslide season in this study area identi�ed June to August as high occurrence months based on the inventory. It was established that antecedent rainfall conditions play an important role in the determination of the groundwater head, and thus the susceptibility of the soil column against shallow landslide failure. In this section an analysis of the seasonal mean seasonal rainfall (MSR) over Wanzhou is conducted over the 40-year period of spatially distributed data from the CMFD gridded observation dataset. Although the landslide season is de�ned as June to August, the months of May to August were considered in the de�nition of seasonal rainfall for this section. The observations in Section 5.1.3 �nds that mean seasonal rainfall conditions were found to represent triggering rainfall conditions, shown in Figure 5.16. This temporal perspective was adapted to assess the antecedent rainfall conditions that in�uence the lateral subsurface �ow in June, while including the potential in�uence of rainfall conditions in August to the subsurface �ow. 78 Part 2: An Analysis of the Present Summer Season Rainfall Section 5.2 Figure 5.17: Average seasonal mean rainfall (MSR) over the study area from May to August 1979 to 2018 over Wanzhou (blue solid line), with maximum MSR in the study area (red dashed line), and the minimum MSR (yellow dashed line) Figure 5.17 depicts the temporal variation of the MSR over the study area. The spatial mean is depicted in blue, while the red and yellow dashed lines represent the spatial maximum and minimum. A signi�cant spatial variation of MSR can be observed for certain periods within the study area, based on the deviation based on the temporal trend of the envelope. An expanded width of the envelope from the spatial mean is observed between the years 1985 and 1987, 2002 and 2004 and consistently from 2007 until 2013. While Figure 5.18 presents the spatial variation between minima and maxima on a temporal scale, it depicts the variation between extremes while essentially comparing the spatial mean with two other points found within the study area. Thus, limiting the general observations of the trend in spatial variability in MSR. Therefore, Figure 5.18 depicts standard deviation as error bars for MSR across the study area to obtain more succinct observations on the spatial variability during this period of observation. Observations derived from Figure 5.17 suggested increased spatial variability during observation visualized with signi�cant width between the minimum and maximum envelope. The information from Figure Figure 5.18 reveals that in the period identi�ed between 1985 to 1987 in Figure 5.17, the spatial variation in 1987 showed signi�cant spatial standard deviation. In the period 2002 to 2004, 2003 and 2004 were years of 79 Chapter 5 Results and Discussion Figure 5.18: Average seasonal mean rainfall (MSR) from May to August 1979 to 2018 over Wanzhou (blue dashed line), with the standard deviation (in error bars). signi�cant spatial variation. Finally, in the period of 2007 to 2013, a more signi�cant spatial variation of MSR was measured across this observation period when compared to the other periods taken into consideration. During this period, a signi�cant spatial variability is suggested in the year 2011. Additionally, the years 1993 and 2000 with higher spatial means, depicted in Figure 5.17 which means drawing nearer to the maximum MSR lines also are periods with higher spatial variability, as suggested by the standard deviation measurement. This analysis suggests that changing conditions in the catchment and the interactions with the atmospheric system have increased the variation in mean seasonal rainfall between May to August. Therefore, providing strong evidence against establishing a one reference scenario MSR value to represent the entire study area. The results of this analysis suggest that establishing a MSR value for each pixel, can better account for the spatial variability observed over the 40-year duration of the CMFD dataset. The MSR values presented in this section were derived from the mean monthly rainfall for the months of May-August. Thus, in order to understand and derive a spatially varying MSR value to represent a reference scenario and account for the variability in rainfall that was observed in across the study area, the spatial distribution of the mean monthly rainfall within the seasonal period as well as the coe�cient of variation of each month was analyzed. 80 Part 2: An Analysis of the Present Summer Season Rainfall Section 5.2 Figure 5.19: Mean monthly rainfall (MMR) showing the average MMR value derived from the period of 1979-2018 for the months of May (A), June (B), July (C), and August (D). 81 Chapter 5 Results and Discussion The MMR for May to August are individually presented in Figure Figure 5.19. The MMR for each month was computed by taking the monthly MMR from 1979-2018. Thus, the MMR in Figure 5.19 represents the average daily rainfall for the month considered within this 40-year observation period. The MMR in May is between 5-6 mm/day, no apparent spatial variation especially within the study area. June shows a range of MMR values between 5.5 mm/day and 6.5 mm/day with a mild decreasing pattern of rainfall diagonally from the northwest to the southwest. The range of daily rainfall values in July are between 5.5 and 7.0 mm/day. The concentration of the highest daily rainfall values in July are observed in the northeast corner of the plots, and outside of the study area. Within the study area, a diagonally increasing pattern of MMR is observed from the southwest to the northeast corners. There is more spatial variability in July, as compared to June and May with the larger magnitude rainfall areas observed in the northwest and southeast corners. The MMR in August appears to have the lowest range of values being between 4.5 mm/day to 5.5 mm/day. There is a similar pattern of increase in rainfall magnitude with that observed in June, daily rainfall increasing from northeast to southwest, except for an observable area of increased magnitude in the northwest corner of the study area. Based on Figure 5.19, June was observed to have the highest distribution of daily rainfall values across all 4 months within the study area. The analysis ofFigure 5.19 depicts the spatial patterns of MMR in each month. A further analysis into the temporal variation of the monthly means was conducted through the calculation of a coe�cient of variation for each pixel. The temporal insight derived from the coe�cient of variation for the MMR of each month is spatially presented in Figure 5.20. The observed temporal variation in the monthly means presented in Figure 5.20, reveals that signi�cant variation can be expected from the means derived for July and August, compared to those derived from May and June. The highest variation is expected for July, where the southwest corner of the area of interest shows a coe�cient of variation between 0.7 and 0.8. This indicates that deviation from the mean monthly rainfall within the 40-year observation period can be 70-80% higher for some years. The relatively low coe�cient of variation in May and June is 0.4, and the visualization in Figure 5.20 indicates that no signi�cant spatial variation is observed. The spatiotemporal analysis of magnitude, pattern and variation of each individual month provides insight on the rainfall characteristics of each month and the potential uncertainty of monthly components when representing MMR on a seasonal time scale. 82 Part 2: An Analysis of the Present Summer Season Rainfall Section 5.2 Figure 5.20: Coe�cients of variation in mean monthly rainfall derived over the period of 1979 to 2018 for the months of May (A), June (B), July (C), and August (D). 83 Chapter 5 Results and Discussion 5.2.2 Analysis of the Mean Seasonal Rainfall Reference Scenario In order to meet the objective of establishing a reference scenario to describe the antecedent rainfall for the landslide season of June to August, the MSR was computed by taking the MMR for the months of May to August for the years 1979 to 2018. Figure 5.21 presents the MSR that represents the reference scenario for antecedent rainfall. The range of MMR magnitudes are between 5.4 mm/day and 6.0 mm/day in the area of interest, with the study area having a range of 5.6 mm/day to 5.9 mm/day. The spatial pattern of rainfall follows a combination of spatial patterns observed in the analysis of the individual month’s MMR. The comparison between the MSR and the monthly derived daily rainfall reveal that on the northeast to southwest diagonal of the area of interest there is an evident increase towards the center of the study area. Meanwhile, the northwest to southeast diagonal of the study area showed an increase in MSR magnitude away from the center. Similar to the spatial pattern of increase observed for July in Figure 5.19, the northwest and southeast corners of the study area contain the magnitudes in the higher range of MSR values. The northeast and southeast corners on the other hand follow the spatial patterns of increase observed in May and August, where the magnitude of MSR values increase diagonally towards the center of the study area. Figure 5.21: Reference scenario derived from mean seasonal rainfall derived from May to August over the period of 1979 to 2018. 84 Part 2: An Analysis of the Present Summer Season Rainfall Section 5.2 The calculation of a coe�cient of variation was employed in the analysis of MSR to account for the temporal uncertainty in deriving this value over the time period of 1979-2018. Figure 5.22 presents the spatial characteristics of the temporal variation in MMR derived on a seasonal time scale. The range of variation is between 0.2 and 0.3, suggesting lower temporal variability in MSR as compared to that of monthly derived daily rainfall, with coe�cients extending until 0.8 for the months of July and August. Furthermore, the variation along the northwest to southeast diagonal within the study area was observed to increase towards the center. An increase in variation was observed along the southwest to northeast diagonal. Figure 5.22: Coe�cient of variation measuring the temporal variation in mean seasonal rainfall from 1979 to 2018. An analysis of the relative deviation between the seasonal mean and each monthly mean was conducted to quantify the ability of the MSR as the reference scenario to represent each monthly MMR over the 40-year period of observation. The relative deviation of the was taken by the di�erence between the monthly MMR and the MSR over the MSR and is reported in percentage. The spatial distribution of this analysis for each month is depicted in Figure 5.23. The range of relative deviation of the MSR is between -20% and +20%. The most signi�cant deviations within the study area are observed in June and August. The MSR underestimates the MMR in June by 5% to 15%, while overestimating the mean of August by 5% to 20%. The spatial characteristics of the relative deviation for these two months show mild di�erence, with August showing 85 Chapter 5 Results and Discussion Figure 5.27: Minimum extreme daily rainfall (A) derived from the monthly block maxima approach, and the relative deviation of the minima each grid cell from the spatial mean of extreme daily rainfall representing relatively dry years (B) from the period of June to August 1979-2018. 92 Part 2: An Analysis of the Present Summer Season Rainfall Section 5.2 Figure 5.28: Skewness coe�cient of the extreme daily rainfall of each grid cell within the study area for 1979 to 2018. Frequency Distribution Modeling of Extreme Daily Rainfall The methodology for modeling the frequency distribution of extreme daily rainfall was described in Section 4.4. The Gumbel distribution model was selected for this research project in order to build on previous knowledge and application of this frequency distribution model in the Wanzhou County area by Xiao et al. [21]. The summary of Gumbel Fit parameters is presented in Table 5.2. A Gumbel �t was estimated for each CMFD pixel within the area of interest, resulting in 120 frequency distribution models. The speci�cations of the minimum, mean and maximum Gumbel �t parameters are also presented in Table 5.2. The spatial variability observed in the analysis of EDR extracted by block maxima approach and the spatial variability of the minimum and maximum EDR results are re�ected in the variability of the tabulated Gumbel Fit parameters. The Goodness-of-�t statistical tests were performed to validate the results of the Gumbel frequency distribution models. Each test is discussed in Section 4.4. The 93 Chapter 5 Results and Discussion Table 5.2: Summary of minimum, mean and maximum Gumbel �t parameters. Gumbel Fit Parameters Minimum Mean Maximum Scale U14.6 17.3 21.7 Beta V29.4 32.9 36.9 Mean `30.3 33.9 38.1 Standard Deviation f18.8 22.2 27.8 con�dence interval applied in this test was 95%. Therefore, the condition of the P-value for each test statistic for each test run was that it would be greater than or equal to 0.05. Also tabulated in Table 5.3 are the results of each test, the correspondent statistic and the p-value, showing the minimum, maximum and mean. The results of the Gumbel distribution �tting to the EDR data across the area of interest provided an adequate model of the frequency distribution. Table 5.3: Summary Goodness-of-�t test statistics with p-values (in parenthesis). Goodness-of-�t Test Minimum Mean Maximum KS Test 0.029 (0.39)0.052 (0.86)0.082 (1.00) AD Test 0.145 (0.53)0.318 (0.91)0.730 (1.00) CVM Test 0.015 (0.44)0.046 (0.89)0.136 (1.00) The resulting parameters of the validated Gumbel distributions were then utilized to estimate daily rainfall corresponding to di�erent return periods, given by Section 4.4. The return period curve was then plotted for each distribution for rainfall corresponding to EDR with return periods between 2 years and 200 years. The return period curves for all the distributions �t within the area of in�uence are plotted in Figure 5.29. The return period curve returning the maximum EDR is plotted in green, while the curve describing the minimum EDR is plotted in red. The individual curves modeled after data from individual time series across the area are plotted in grey. The spatial variability of EDR for di�erent return periods is observed by the density of the grey plotted Gumbel �t lines between the red lower limit and the green upper limit. The divergence of estimation of rainfall as the return period increases also exponentially increases. The spatial variability observed in the return period estimated rainfall re�ects the variability of extreme rainfall in the observations based on the monthly maxima. These results highlight a signi�cant application of gridded precipitation datasets with high spatial resolution to address the characteristics of spatial variability in extreme rainfall. The extreme daily rainfall at higher return periods in Figure 5.29 have a signi�cantly lower values than the maximum extreme daily rainfalls presented in 94 Part 2: An Analysis of the Present Summer Season Rainfall Section 5.2 Figure 5.29: Return period curves derived from Gumbel �ttings in all pixels across the study area (grey lines) with the relationship curves for the mean extreme daily rainfall (blue), the maximum (green), and the minimum (red). Figure 5.26. This is derived from the tendency of the Gumbel distribution to yield the the smallest possible rainfall value in comparison to other extreme value distributions [80]. 5.2.4 An Analysis of the Extreme Daily Rainfall Reference Scenario The objective of modeling extreme daily rainfall in this research project was to establish a spatially varying reference scenario for extreme daily rainfall at di�erent return periods. The signi�cant results of this methodology are presented in this section. The reference scenarios to assess the future of shallow landslide susceptibility was created to give spatial representation to the rainfall corresponding to return periods of 10, 20, 50 and 100 years. The spatial distribution of EDR corresponding to di�erent return periods using the Gumbel distribution parameters was utilized to assess the spatial characteristics of EDR corresponding to di�erent probabilities of occurrence across the area of interest. The spatial distribution of EDR for di�erent return periods of 10, 20 , 50 and 100 years are mapped in Figure 5.30. 95 Chapter 5 Results and Discussion Figure 5.30: Comparison of the spatial distribution in extreme daily rainfall corresponding to return periods of 10, 20, 50 and 100 years (left to right, top to bottom) calculated through the Gumbel �ttings. 96 Part 3: Projected Summer Rainfall Under Climate Change Section 5.3 A signi�cant observation of the rainfall return levels in Figure 5.30 an increasing side of the large magnitude EDR at the center of the catchment. This clustering of EDR from the center of the catchment grows in an area for EDR between 10-year to 20-year return periods. The di�erences in magnitude between the areas visualized in Figure 5.30 are de�ned by di�erences of 10 mm/day. This area of clustered EDR proceeds to decrease from 20-year to 50-year EDR estimates, and alternatively increasing to an area even larger than the 20-year return level cluster area for 100-year return levels. The area in the center of the study area is of concern, suggesting with a 100-year return period, a growing area exposed to risk of 120 mm to 130 mm rainfall magnitudes emerges. The spatial pattern for this cluster of EDR was observed in the analysis of the maximum EDR values Figure 5.30, this area was concentrated in the center to northeast of the study area. Although Figure 5.30 identi�es a growing area of risk in the center of the study area, the west border of the area of interest also reveals a signi�cant trend of increasing EDR with return levels greater than those modeled within the study area. The magnitudes of these EDR values are between 130 mm to 150 mm and can represent signi�cant triggering-rainfall conditions that could induce shallow landslides. This risk area lies within a valley between mountain ranges reaching peak heights of 1640 m on both sides, shown in Figure 1.1. The location of increased precipitation suggests an orographic enhancement of precipitation could be driving increased EDR magnitudes during the summer season. These estimates are outside of the boundaries of the study area, and the spatial uncertainty that accompanies the the CMFD dataset spatial interpolation, integration of satellite estimates and reanalysis results prompts the recommendation of the possibility that the EDR on the western border potentially in�uences to instability stability of the western region of the study area. 5.3 Part 3: Projected Summer Rainfall Under Climate Change This section of the results presents the implementation of Section 4.5 to derive projections and conduct a climate change analysis. A model bias correction of the Regional climate model outputs through the Quantile Delta Mapping (QDM) method was conducted. The climate signals of the daily rainfall and extreme daily rainfall were separately corrected with transfer functions derived through an empirical CDF, and a parametric Gumbel �t CDF, respectively. A multi-model ensemble of 4 climate model outputs was considered. This combination comprised of the HadGEM-ES and MPI-ESM Global climate models (GCM) providing the boundary layer conditions for dynamic downscaling by the RegCM4 and REMO Regional climate models (RCM). The 97 Chapter 5 Results and Discussion spatio-temporal projections and uncertainties derived from the mean ensemble were analyzed for the scenarios de�ned by Table 4.2. 5.3.1 Validation of Bias Correction by �antile Delta Methods Bias correction was applied to the climate model outputs to address the systematic errors present in the results. A bias correction on all ensemble member outputs were performed separately for daily precipitation and extreme daily rainfall, as described in Section 4.5.2. The cross-validation procedure was performed to assess the performance of the empirical QDM on daily rainfall results within the reference period of 1979-2018, as described in Section 4.5.3. A validation procedure for the bias-corrected extreme daily rainfall (EDR) was performed, following Section 4.5.4, and compared the observed EDR over the reference period. Cross-Validation of Bias-corrected Daily Rainfall The results of the bias correction methodology for daily rainfall, described in Section 4.5.2, was applied to the historical model results from 1979-2005 combined with the model projections from 2005-2018 of each ensemble member. The bias-corrected results over this combined period representing the historical and projection outputs were then compared to the CMFD observations through a cross-validation procedure. This section presents the results of the cross-validation for bias-corrected RCM outputs for daily rainfall, following the methodology in Section 4.5.3.Figure 4.5 provides an illustrative box plot comparison between the cross-validation performance of the bias-correction for daily rainfall as measured by the mean average error (MAE). This measure for cross-validation was employed to measure the accuracy of the bias-correction. The MAE performance of the bias-corrected ensemble members across the study area ranges between 8.6 mm/day to 8.8 mm/day. The interquartile ranges of the MAE for all four ensemble members showed little variation with ranges between 8.5 mm/day to 8.9 mm/day. Furthermore, the cross-validation performance measured by root-mean squared error (RMSE) was also calculated and is shown in Figure 5.32. The RMSE gives insight into the average magnitude of error and penalizes larger deviations with relatively higher weight The range of mean RMSE across all ensemble members is from 15.5 mm/day to 16 mm/day. The interquartile ranges are from 15 mm/day to 16.5 mm/day. The minimum box plot whiskers are equivalent across the ensemble with values below 14.5 mm/day, the maximum whiskers for the REMO RCM results are observably larger than the RegCM4 maxima. Although this may indicate better performance of the RegCM4 members on the extreme, the interquartile range of the REMO RCM models is evidently smaller. The analysis of MAE and RMSE performance in the cross-validation 98 Part 3: Projected Summer Rainfall Under Climate Change Section 5.3 Figure 5.31: Box plot comparison of cross-validation performance of the daily rainfall bias correction measured by mean average error (MAE) for each grid point results of the four GCM-RCM ensemble member combinations. presented in Figure 5.31 and Figure 5.32 provided to no conclusive evidence suggesting one GCM-RCM combination outperforms the other. Validation of Bias-corrected Extreme Daily Rainfall Extreme daily rainfall bias correction, described by the methodology in Section 4.5.2 was performed for all ensemble member outputs over a period incorporating the the historical model results from 1979-2005 combined with the model projections from 2005-2018 of each ensemble member. The parametric quantile mapping bias correction was performed with a Gumbel distribution �t, given in Section 4.4 to the data to yield each ensemble member’s output for extreme rainfall. The reference period observations from the CMFD dataset from 1979-2018 ordered and used to validate the ordered and bias-corrected extreme rainfall results of each ensemble member, as described in Section 4.5.4. The ordered statistics of the datasets were utilized to measure the performance of the bias-corrected historical ensemble member results during reference period. The performance was �rst measured using the Mean Absolute Error (MAE) to give insight into the uncertainty in the corrected accuracy. The comparison of the MAE performance is presented in Figure 5.33. The range MAE for the EDR bias correction varies from 4 99 Chapter 5 Results and Discussion Figure 5.32: Box plot comparison of cross-validation performance of the daily rainfall bias correction measured by root-mean squared error (RMSE) for each grid point results of the four GCM-RCM ensemble member combinations. mm/day to 5.5 mm/day. The performance of the MPI RCM results is evidently better than the HadGEM RCM outputs in Figure 5.33. The interquartile range of MPI-driven outputs are between 2 mm/day and 4 mm/day with maximum error measured reaching 5 mm/day. The performance of the HadGEM bias-corrected outputs are between 4 mm/day to 6 mm/day with a maximum 7.5 mm/day. This comparison of MAE performance validated against the reference period observations suggest that the boundary conditions provided by the MPI GCM provide a more accurate perspective on the extreme daily rainfall. Similar to the analysis of performance in the daily rainfall bias correction results, the bias-corrected EDR performance was also measured through root-mean squared error (RMSE). The results of the performance validation of the bias-corrected results versus the observations during the reference period are presented in Figure 5.34. The results of the comparison of RMSE on the ordered statistics suggest that the MPI-driven bias corrections outperform the HadGEM-driven outputs. This is supported by the interquartile range of the MPI models being between 4 mm/ day to 8 mm/day, compared to the HadGEM outputs having a range of 5 mm/day to 10 mm/day. Furthermore, the outliers beyond the maximum whiskers are signi�cantly less and lower in RMSE magnitude for MPI-driven models as compared to HadGEM-driven models. The analysis of RMSE in Figure 5.33 and Figure 5.34 give reasonable evidence to suggest that the 100 Part 3: Projected Summer Rainfall Under Climate Change Section 5.3 Figure 5.33: Comparison of validation performance of the extreme daily rainfall bias correction measured by mean average error (MAE) for each grid point results of the four GCM-RCM ensemble member combinations. bias-corrected MPI-driven RCM models perform better than the HadGEM-driven RCM models in capturing extreme daily rainfall. The performance of the ensemble was initially measured with MAE and RMSE to initially assess the accuracy and magnitude of deviation in capturing EDR observed over the reference period. It was further determined that since the bias-correction for EDR was parametric and performed through quantile mapping correction of �tted Gumbel distribution models on the historical and observed data sets, an assessment of the performance using the Pearson correlation coe�cient between the magnitude of the EDR observed and the modeled EDR was necessary to further understand the reliability of each ensemble member. The comparison of the box plots capturing the distribution of Pearson correlation coe�cients for each ensemble member across the study area is presented in Figure 5.35 The interquartile ranges of the MPI-driven results are between 0.96 and 0.99, indicating very high correlation between the magnitude of EDR modeled and those observed. The interquartile ranges of the HadGEM-driven models are between 0.93 and 0.98, also indicating good performance in capturing EDR. Although the MPI driven results generally perform better than the HadGEM results, as observed in Figure 5.33 and Figure 5.34, the analysis of Pearson correlation coe�cients support the notion that the ensemble member combinations perform adequately in capturing EDR. 101 Chapter 5 Results and Discussion between 0.9 to 1.0. Within the western region of increase is a pocket area with an increase in ⇠⇠⇢⇡',) for this scenario ranging from 1.0 to 1.2. The eastern region of the study areas shows a consistent CCF between 1.0 and 1.2. Although a small area in the southeast region of Wanzhou is projected to have a range of CCFs below 1.0, this region reduces in area as the return period increases from 5 years to 50 years, and is not evident in the values for 100-year events, ⇠⇠⇢⇡',100. The spatial distribution of ⇠⇠⇢⇡',) in the Mid 21st Century indicates that more areas will be subject to a decrease in magnitude when considering extreme daily rainfall with a return period of 5 to 10 years. Though, once higher return period ) in ⇠⇠⇢⇡',) are considered, more areas will experience a mild increase in magnitude. This is evident in the projections ⇠⇠⇢⇡',50 and ⇠⇠⇢⇡',100 year extreme daily rainfall. The spatial distribution of the projected values of ⇠⇠⇢⇡',) in the Late 21st Century projections, were presented in Figure 5.38. The ensemble mean projections for extreme daily rainfall corresponding to return periods of 5, 10, 50 and 100 years were illustrated. Two characteristic regions, categorized by magnitude of ensemble ⇠⇠⇢⇡',) are visible across return periods in Figure 5.38. The �rst region represents areas with projected ⇠⇠⇢⇡',) between 1.2 and 1.4, predominantly covering the extent of the Wanzhou County area. The second region projects higher magnitudes of extreme daily rainfall with ⇠⇠⇢⇡',) between 1.4 and 1.6 located around the study area and at the borders. The areas projecting higher extreme daily rainfall ⇠⇠⇢⇡',) are evident in the southwest region for 5-year return period values, ⇠⇠⇢⇡',5 , and increase to encircle the study area at the borers as the spatial distribution of return periods increase from 10, 50 to 100 years. The central region of the Wanzhou County area and the north east sees a consistent pattern of ensemble ⇠⇠⇢⇡',) projections ranging between 1.2 to 1.4 for EDR, while the southeastern region sees an increase in projected CCF for EDR of 1.4 to 1.6. The potential in�uence of orography in enhancing extreme daily rainfall under is evident in the Late 21st Century scenario projections. The risk area identi�ed in Figure 5.30 has shifted northwest towards the northern mountain range at the boundary of Wanzhou County, illustrated in Figure 1.1. Several factors under climate change conditions can be attributed to the orogaphically enhanced extreme daily rainfall. Sandvik et al. [81]�nd that a change in precipitation phases result in signi�cant increases in rainfall with increasing temperature and elevation in a study on historical orographically enhanced extreme precipitation events in Norway. Napoli et al. [82] see a historical di�erence between lowland and highland distributions of annual precipitation from the mid 20th century to the 1980’s over the European Alpine region. The interdecadal increase was attributed to increased anthropogenic aerosol load, thus a causal link between antrhopogenic activity and the orographic enhancement of precipitation is suggested. 108 Part 3: Projected Summer Rainfall Under Climate Change Section 5.3 Figure 5.38: Mid 21st Century scenario projections of extreme daily rainfall climate change factors ( ⇠⇠⇢⇡',) ) for return periods ( ) ) of 5, 10, 50 and 100 years derived from the ensemble mean projections covering the months of June to July. 109 Chapter 5 Results and Discussion Figure 5.39: Late 21st Century scenario projections of extreme daily rainfall climate change factors ( ⇠⇠⇢⇡',) ) for return periods ( ) ) of 5, 10, 50 and 100 years derived from the ensemble mean projections covering the months of June to July. 110 Part 3: Projected Summer Rainfall Under Climate Change Section 5.3 The uncertainty in the mean ensemble projections for the future climate change scenarios in the Mid and Late 21st Centuries were assessed and measured using the coe�cient of variation in ⇠⇠⇢⇡',) derived from the standard deviation of the ensemble. The coe�cient of variation was calculated for each return period and the spatial distribution of this variation was analyzed. This assessment considers the uncertainty and variation in projected extreme daily rainfall with return periods of 5, 10, 50 and 100 year. The spatial distribution for the coe�cient of variation for the Mid 21st Century ensemble projections in Figure 5.40. The range of variation in ( ⇠⇠⇢⇡',) under the projections for this scenario is below 25% within the Wanzhou County area. There is an evident growth in uncertainty in the northern borders of the study area as the return period increases from 10 years to 100 years. Though the ensemble variation in this area is between 15% to 25%, a majority of the study area remains within the range of 10% to 15%. The least uncertainty in projections is found across the study area at lower return periods of 5 and 10 years. The analysis of the spatial distribution in coe�cients of variation in the Late 21st Century scenario is shown in Figure 5.41 The range of the variation of this scenario is between 0% and 30%. The variation is similar to the range observed in Mid 21st Century projections. The spatial distribution of the ensemble variation in the Late 21st Century di�ers from the Mid 21st Century as the return period increases. A larger area of increased variation within the study area was observed in the Mid 21st Century as the return period increased from 10, 50 to 100 years, the variation for within the study area decreases in the Late 21st Century projections. This trend is particularly evident in the eastern region of the Wanzhou County area. The coe�cient of variation reduces from being within a range of 15% to 25%, corresponding to 10-year EDR, to a range of 10% to 20% in considering 50 and 100-year EDR. These results of this analysis indicates that there is less uncertainty in the projection of ⇠⇠⇢⇡',) applicable to low frequency and high magnitude EDR in the Late 21st Century. 111 Chapter 5 Results and Discussion Figure 5.40: Coe�cients of variation for Mid 21st Century scenario projection ensemble projections of climate change factors ( ⇠⇠⇢⇡',) ) with return periods ( ) )of 5, 10, 50 and 100 years derived from the ensemble mean projections covering the months of June to July. 112 Part 3: Projected Summer Rainfall Under Climate Change Section 5.3 Figure 5.41: Coe�cients of variation for Late 21st Century scenario projection ensemble projections of climate change factors ( ⇠⇠⇢⇡',) ) with return periods ) of 5, 10, 50 and 100 years derived from the ensemble mean projections covering the months of June to July. 113 6Conclusions and Recommendations The �nal chapter of this study comprises of conclusions and recommendations derived from the limitations and results presented. This chapter aims to connect the conclusions and limitations in the methodology to provide recommendations in developing the methodological framework. The intention is to open insight for research lines that could result in the improvement of the framework with contemporary literature illustrating promising solutions for future lines of research. The primary objective of this research project was to investigate the in�uence of climate change on rainfall conditions that could trigger shallow landslides. In order to investigate this primary research question, this study established a methodology to determine historical triggering rainfall conditions, assess the present rainfall conditions in the study area, and utilize climate model outputs to derive projections in the future. This study presented a procedure to create scenarios from climate model outputs that could be integrated in slope stability model inputs. Through the presentation of results and the identi�cation of sources of uncertainties, a viable methodology that can link climate change models as input in slope susceptibility assessment studies was presented. The methodological framework developed in Figure 4.1 consisted of three main tasks to assess the in�uence of climate change on shallow landslide triggering rainfall conditions in Wanzhou County, China. The �rst task was to reconstruct the triggering rainfall conditions from 1995 to 2005. Second, to establish a reference scenario based on a present analysis of mean seasonal rainfall and extreme daily rainfall. Third, to analyze an ensemble of bias-corrected Regional Climate models (RCM) outputs to produce projections in the Mid 21st Century (2021-2060) and the Late 21st Century (2061-2100). The Reconstruction of Triggering Rainfall Conditions The �rst part of this research project was centered around reconstructing event rainfall and antecedent rainfall conditions for historical shallow landslides. The �ndings from the attempt to reconstruct the rainfall conditions in the inventory indicated that signi�- cant temporal uncertainty exists in the record dates of shallow landslide occurrences. This was primarily observed in the analysis of landslides from June to August that occurred during periods with zero antecedent recharge and no signi�cant event rainfall. The presence of signi�cant temporal uncertainty was further assessed through a sensitivity analysis. The procedure involved the detection of maximum daily rainfall 115 Chapter 6 Conclusions and Recommendations prior to the record date that could possibly represent the event rainfall, and indicate the actual landslide date. The �ndings from this analysis revealed that 88% to 96% of the inventory could have occurred up to 7 days prior to the inventory record date. It was determined through this sensitivity analysis that a 7-day window prior to the record dates carried maximum uncertainty. The inventory record date is not the lone cause for uncertainty in this analysis. The CMFD gridded precipitation data set used to reconstruct the rainfall conditions is limited by its spatial resolution of 0.1 > or approximately 11 km 2 . Another limitation of this analysis is the temporal resolution of daily precipitation selected in this study. While uncertainties in the inventory and the limitations of the CMFD precipitation data set were identi�ed in this portion of the research project, the reconstruction of the event rainfall at a daily time scale was not successfully performed. This was evident in an analysis of the 2D density plot representation of antecedent rainfall %0 and event rainfall %4 . While the clusters formed in the data suggested groups of landslides occurring with a range of %0 values between 3 and 6 mm/day and %4 with values between 10 and 30 mm/day, the uncertainty in the record date makes it di�cult to establish conclusive insight from these ranges. The temporal uncertainty had less in�uence over reconstructed antecedent rainfall over a 30-day period for the months of May to July. In an assessment of the reconstructed antecedent rainfall with a duration of 10, 15 and 30 days, it was found that reconstructing antecedent rainfall with a 15 and 30-day duration proved to be less sensitive to temporal uncertainty. A Kolmogorov-Smirnov test indicated statistical signi�cance in the ability of antecedent rainfall with a 15 and 30-day duration to capture the frequency distribution of the triggering antecedent rainfall distributions with a 7-day window of uncertainty. Therefore, a reconstruction of 30-day antecedent rainfall based on inventory dates adjusted for a 7-day period of uncertainty in the maximum event rainfall was performed. A comparison of the reconstructed 30-day antecedent rainfall versus the CMFD observations of daily rainfall across the study area was conducted in Figure 5.16. It was determined that the distribution of the reconstructed 30-day antecedent rainfall corresponding to shallow landslide occurrences in Wanzhou had a mean value that was statistically similar to that of the mean seasonal rainfall for May to August. Therefore it was concluded that the mean seasonal rainfall can adequately represent the antecedent rainfall that can trigger shallow landslides in Wanzhou County. This study determined the approach of implementing a time-shifted window of uncertainty to detect signi�cant volumes of daily rainfall on the date of the landslides would not address the temporal uncertainty in the record date, and increase the uncertainty in the data set. It is therefore proposed that an exploration of remote sensing techniques to incorporate satellite-derived data and aerial photographs. This can speci�cally be 116 Conclusions and Recommendations Chapter 6 applied to reconstruct the occurrence of shallow water during the landslide seasons of 1995-2005. A study by Miura & Nagai [83] investigated landslide mapping of landslides in 2017 in the northern area of Kyushu, Japan using Advanced Himawari Imager (AHI)- derived normalized di�erence vegetation index (NDVI) time series data. Their results demonstrated that moderate and low spatial resolution data have the potential to successfully detect landslides. While the scale of this research provides a regional overview on shallow landslide triggering rainfall occurrences, the adoption of this scale to the reconstruction of triggering event rainfall limits the level of spatial certainty of identifying event rainfall. Shallow landslides can occur due to intense local rainstorms on signi�cantly smaller spatial scales. Therefore matching the spatial scale of the landslides with appropriate precipitation data is essential to reducing uncertainty in determining the magnitude of the triggering rainfall conditions. Tseng et al. [84]�nds that high-resolution radarderived quantitative precipitation estimates can be adopted to obtain better spatiotemporal precipitation patterns with a spatial resolution of 1 km, in contrast to the CMFD data resolution of approximately 11 km. Although the applicability of this suggestion is greatly limited by the quality and availability of radar-derived rainfall data, this study highlights the importance and potential of incorporating such data sets in the process of inventory reconstruction. This study estimated the antecedent recharge over the entire catchment by the application of temporally-varying recharge parameters obtained from water balance calculations performed with the EasyBal hydrological model for one point in the study area [65]. The spatial variation of climate and hydrologic characteristics across the study area limits the insight from the estimation of antecedent recharge applied in this research. Kim et al. [26] recommend detailed regional scale hydrological model results integrated to capture more precise in�ltration calculations, and gain a better understanding of soil moisture conditions. This insight is critical in the reconstruction of triggering rainfall conditions, and in predicting the watershed response to under a changing climate. Huang et al. [38] uses a variable in�ltration capacity (VIC) hydrology model and regional climate model (RCM) outputs to assess the hydrological responses of the Upper Yangtze River Basin to climate change from 2020 to 2050. The integration of hydrological results, and a thorough comparison of historical watershed responses to the future responses will add value to advancing methods in the reconstruction and projections of landslide triggering rainfall conditions. The integration of the impact of climate change on the hydrological cycle is crucial in incorporating the in�uence of temperature and precipitation on recharge and the groundwater table. 117 11. Medina, V., Hürlimann, M., Guo, Z., Lloret, A. & Vaunat, J. Fast physically-based model for rainfall-induced landslide susceptibility assessment at regional scale. CATENA 201, 105213. ����: 0341-8162. https://www.sciencedirect.com/science/article/pii/S0341816221000722 (2021) (June 1, 2021) (cit. on pp. 2,3). 12. Scheidl, C. et al. The in�uence of climate change and canopy disturbances on landslide susceptibility in headwater catchments. Science of The Total Environment 742, 140588. ����: 0048-9697. https://www.sciencedirect.com/science/article/pii/S0048969720341103 (2021) (Nov. 10, 2020) (cit. on pp. 2,4,5,17). 13. Aristizábal, E., Vélez, J. I., Martínez, H. E. & Jaboyedo�, M. SHIA_Landslide: a distributed conceptual and physically based model to forecast the temporal and spatial occurrence of shallow landslides triggered by rainfall in tropical and mountainous basins. Landslides 13, 497–517. ����: 1612-5118. https://doi.org/10.1007/s10346-015-0580-7 (2021) (June 1, 2016) (cit. on p. 3). 14. Segoni, S., Piciullo, L. & Gariano, S. L. A review of the recent literature on rainfall thresholds for landslide occurrence. Landslides 15, 1483–1501. ����: 1612-510X, 1612-5118. http: //link.springer.com/10.1007/s10346-018-0966-4 (2021) (Aug. 2018) (cit. on pp. 3,39). 15. Polemio, M. & Petrucci, O. Occurrence of landslide events and the role of climate in the twentieth century in Calabria, southern Italy. QUARTERLY JOURNAL OF ENGINEERING GEOLOGY & HYDROGEOLOGY 43. https://trid.trb.org/view/1100062 (2021) (Nov. 2010) (cit. on p. 4). 16. Donat, M. G., Lowry, A. L., Alexander, L. V., O’Gorman, P. A. & Maher, N. More extreme precipitation in the world’s dry and wet regions. Nature Climate Change 6. Number: 5 Publisher: Nature Publishing Group, 508–513. ����: 1758-6798. https://www.nature.com/ articles/nclimate2941 (2021) (May 2016) (cit. on p. 4). 17. Brönnimann, S. et al. Changing seasonality of moderate and extreme precipitation events in the Alps. Natural Hazards and Earth System Sciences 18. Publisher: Copernicus GmbH, 2047–2056. ����: 1561-8633. https://nhess.copernicus.org/articles/18/2047/2018/ (2021) (July 27, 2018) (cit. on p. 5). 18. Liu, Y., Yin, K., Chen, L., Wang, W. & Liu, Y. A community-based disaster risk reduction system in Wanzhou, China. International Journal of Disaster Risk Reduction 19, 379–389. ����: 2212-4209. https://www.sciencedirect.com/science/article/pii/S2212420916303405 (2021) (Oct. 1, 2016) (cit. on pp. 6,20). 19. Xiao, T., Segoni, S., Chen, L., Yin, K. & Casagli, N. A step beyond landslide susceptibility maps: a simple method to investigate and explain the di�erent outcomes obtained by di�erent approaches. Landslides 17, 627–640. ����: 1612-5118. https://doi.org/10.1007/ s10346-019-01299-0 (2021) (Mar. 1, 2020) (cit. on pp. 6,19). 20. Huang, F., Yin, K., Huang, J., Gui, L. & Wang, P. Landslide susceptibility mapping based on self-organizing-map network and extreme learning machine. Engineering Geology 223, 11– 22. ����: 0013-7952. https://www.sciencedirect.com/science/article/pii/S0013795216303143 (2021) (June 7, 2017) (cit. on pp. 6,20–22). 124 21. Xiao, L., Wang, J., Zhu, Y. & Zhang, J. Quantitative Risk Analysis of a Rainfall-Induced Complex Landslide in Wanzhou County, Three Gorges Reservoir, China. International Journal of Disaster Risk Science 11, 347–363. ����: 2192-6395. https://doi.org/10.1007/s13753020-00257-y (2021) (June 1, 2020) (cit. on pp. 6,20,22,23,41,48,93). 22. Coles, S. An Introduction to Statistical Modeling of Extreme Values ����: 978-1-85233-459-8. https://www.springer.com/gp/book/9781852334598 (2021) (Springer-Verlag, London, 2001) (cit. on pp. 11,41). 23. Crozier, M. Landslides: Causes, Consequences and Environment. (Routledge, London, 1989) (cit. on p. 12). 24. Hong, M., Kim, J. & Jeong, S. Rainfall intensity-duration thresholds for landslide prediction in South Korea by considering the e�ects of antecedent rainfall. Landslides 15, 523–534. ����: 1612-510X, 1612-5118. http://link.springer.com/10.1007/s10346-017-0892-x (2021) (Mar. 2018) (cit. on p. 12). 25. Khan, Y. A., Lateh, H., Baten, M. A. & Kamil, A. A. Critical antecedent rainfall conditions for shallow landslides in Chittagong City of Bangladesh. Environmental Earth Sciences 67, 97–106. ����: 1866-6280, 1866-6299. http://link.springer.com/10.1007/s12665-011-1483-0 (2021) (Sept. 2012) (cit. on p. 12). 26. Kim, S. W. et al. E�ect of antecedent rainfall conditions and their variations on shallow landslide-triggering rainfall thresholds in South Korea. Landslides 18, 569–582. ����: 1612510X, 1612-5118. http://link.springer.com/10.1007/s10346-020-01505-4 (2021) (Feb. 2021) (cit. on pp. 12,39,48,117). 27. Ma, T., Li, C., Lu, Z. & Wang, B. An e�ective antecedent precipitation model derived from the power-law relationship between landslide occurrence and rainfall level. Geomorphology 216, 187–192. ����: 0169555X. https://linkinghub.elsevier.com/retrieve/pii/ S0169555X14001676 (2021) (July 2014) (cit. on p. 12). 28. Marques, R., Zêzere, J., Trigo, R., Gaspar, J. & Trigo, I. Rainfall patterns and critical values associated with landslides in Povoação County (São Miguel Island, Azores): relationships with the North Atlantic Oscillation. Hydrological Processes 22, 478–494. ����: 08856087, 10991085. http://doi.wiley.com/10.1002/hyp.6879 (2021) (Feb. 15, 2008) (cit. on p. 12). 29. Mathew, J., Babu, D. G., Kundu, S., Kumar, K. V. & Pant, C. C. Integrating intensity– duration-based rainfall threshold and antecedent rainfall-based probability estimate towards generating early warning for rainfall-induced landslides in parts of the Garhwal Himalaya, India. Landslides 11, 575–588. ����: 1612-510X, 1612-5118. http://link.springer. com/10.1007/s10346-013-0408-2 (2021) (Aug. 2014) (cit. on pp. 12,13,39). 30. Tang, D., Li, D. - Q. & Cao, Z. - J. Slope stability analysis in the Three Gorges Reservoir Area considering e�ect of antecedent rainfall. Georisk: Assessment and Management of Risk for Engineered Systems and Geohazards 11, 161–172. ����: 1749-9518, 1749-9526. https://www.tandfonline.com/doi/full/10.1080/17499518.2016.1193205 (2021) (Apr. 3, 2017) (cit. on pp. 12,15,39). 125 31. Tu, X., Kwong, A., Dai, F., Tham, L. & Min, H. Field monitoring of rainfall in�ltration in a loess slope and analysis of failure mechanism of rainfall-induced landslides. Engineering Geology 105, 134–150. ����: 00137952. https://linkinghub.elsevier.com/retrieve/pii/ S0013795208003153 (2021) (Apr. 2009) (cit. on pp. 12,13). 32. Segoni, S., Pappa�co, G., Luti, T. & Catani, F. Landslide susceptibility assessment in complex geological settings: sensitivity to geological information and insights on its parameterization. Landslides 17, 2443–2453. ����: 1612-5118. https://doi.org/10.1007/ s10346-019-01340-2 (2021) (Oct. 1, 2020) (cit. on p. 12). 33. Dahal, R. K. & Hasegawa, S. Representative rainfall thresholds for landslides in the Nepal Himalaya. Geomorphology 100, 429–443. ����: 0169-555X. https://www.sciencedirect.com/ science/article/pii/S0169555X08000172 (2021) (Aug. 15, 2008) (cit. on pp. 13,14,39). 34. Alvioli, M. et al. Implications of climate change on landslide hazard in Central Italy. Science of The Total Environment 630, 1528–1543. ����: 00489697. https://linkinghub.elsevier.com/ retrieve/pii/S0048969718307150 (2021) (July 2018) (cit. on pp. 16,17). 35. Song, Y. et al. Landslide Susceptibility Mapping Based on Weighted Gradient Boosting Decision Tree in Wanzhou Section of the Three Gorges Reservoir Area (China). ISPRS International Journal of Geo-Information 8. Number: 1 Publisher: Multidisciplinary Digital Publishing Institute, 4. https://www.mdpi.com/2220-9964/8/1/4 (2021) (Jan. 2019) (cit. on pp. 20,21). 36. Xiao, T., Yin, K., Yao, T. & Liu, S. Spatial prediction of landslide susceptibility using GISbased statistical and machine learning models in Wanzhou County, Three Gorges Reservoir, China. Acta Geochimica 38, 654–669. ����: 2365-7499. https://doi.org/10.1007/s11631-01900341-1 (2021) (Oct. 1, 2019) (cit. on p. 21). 37. Tong, Y. et al. Bias correction of temperature and precipitation over China for RCM simulations using the QM and QDM methods. Climate Dynamics. ����: 1432-0894. https: //doi.org/10.1007/s00382-020-05447-4 (2021) (Sept. 8, 2020) (cit. on pp. 27,46,47,107,120). 38. Huang, Y. et al. Hydrological projections in the upper reaches of the Yangtze River Basin from 2020 to 2050. Scienti�c Reports 11. Bandiera_abtest: a Cc_license_type: cc_by Cg_type: Nature Research Journals Number: 1 Primary_atype: Research Publisher: Nature Publishing Group Subject_term: Climate change;Hydrology Subject_term_id: climatechange;hydrology, 9720. ����: 2045-2322. https://www.nature.com/articles/s41598-02188135-5 (2021) (May 6, 2021) (cit. on pp. 27–29,31,46,48,103,117,118,120). 39. Yang, Y., Tang, J., Xiong, Z., Wang, S. & Yuan, J. An intercomparison of multiple statistical downscaling methods for daily precipitation and temperature over China: present climate evaluations. Climate Dynamics 53, 4629–4649. ����: 1432-0894. https://doi.org/10.1007/ s00382-019-04809-x (2021) (Oct. 1, 2019) (cit. on pp. 27,46,121). 40. Gao, X., Xu, Y., Zhao, Z., Pal, J. S. & Giorgi, F. On the role of resolution and topography in the simulation of East Asia precipitation. Theoretical and Applied Climatology 86, 173–185. ����: 1434-4483. https://doi.org/10.1007/s00704-005-0214-4 (2021) (Sept. 1, 2006) (cit. on pp. 27,45). 126 41. Gao, X. et al. Performance of RegCM4 over major river basins in China. Advances in Atmospheric Sciences 34, 441–455. ����: 1861-9533. https://doi.org/10.1007/s00376-0166179-7 (2021) (Apr. 1, 2017) (cit. on pp. 29,30,121). 42. Liu, J., Du, J., Yang, Y. & Wang, Y. Evaluating extreme precipitation estimations based on the GPM IMERG products over the Yangtze River Basin, China. Geomatics, Natural Hazards and Risk 11. Publisher: Taylor & Francis _eprint: https://doi.org/10.1080/19475705.2020.1734103, 601–618. ����: 1947-5705. https://doi.org/10.1080/19475705.2020.1734103 (2021) (Jan. 1, 2020) (cit. on pp. 30,34). 43. Jingwei, X. et al. The Assessment of Surface Air Temperature and Precipitation Simulated by Regional Climate Model REMO over China. CLIMATE CHANGE RESEARCH 12. Number: 4, 286–293. ����: 1673-1719. http://www.climatechange.cn (2021) (July 2016) (cit. on p. 31). 44. Xu, J. et al. Downstream e�ect of Hengduan Mountains on East China in the REMO regional climate model. Theoretical and Applied Climatology. Publisher: SPRINGER WIEN. ����: 0177-798X. http://doi.org/10.1007/s00704-018-2721-0 (2021) (Dec. 14, 2018) (cit. on p. 31). 45. Gu, H., Yu, Z., Yang, C. & Ju, Q. Projected Changes in Hydrological Extremes in the Yangtze River Basin with an Ensemble of Regional Climate Simulations. Water 10. Number: 9 Publisher: Multidisciplinary Digital Publishing Institute, 1279. https://www.mdpi.com/ 2073-4441/10/9/1279 (2021) (Sept. 2018) (cit. on p. 31). 46. Gu, H. et al. Assessing CMIP5 general circulation model simulations of precipitation and temperature over China. International Journal of Climatology 35, 2431–2440. ����: 1097-0088. https://rmets.onlinelibrary.wiley.com/doi/abs/10.1002/joc.4152 (2021) (2015) (cit. on p. 32). 47. He, J. et al. The �rst high-resolution meteorological forcing dataset for land process studies over China. Scienti�c Data 7. Bandiera_abtest: a Cc_license_type: cc_publicdomain Cg_type: Nature Research Journals Number: 1 Primary_atype: Research Publisher: Nature Publishing Group Subject_term: Environmental sciences;Hydrology Subject_term_id: environmental-sciences;hydrology, 25. ����: 2052-4463. https://www.nature.com/articles/ s41597-020-0369-y (2021) (Jan. 21, 2020) (cit. on pp. 33,35). 48. Jacob, D. et al. Assessing the Transferability of the Regional Climate Model REMO to Di�erent COordinated Regional Climate Downscaling EXperiment (CORDEX) Regions. Atmosphere 3. Number: 1 Publisher: Molecular Diversity Preservation International, 181– 199. https://www.mdpi.com/2073-4433/3/1/181 (2021) (Mar. 2012) (cit. on pp. 33,38). 49. Zhe, L., DaWen, Y. & Hong, Y. Multi-scale evaluation of high-resolution multi-sensor blended global precipitation products over the Yangtze River. Journal of Hydrology (Amsterdam) 500. Publisher: Elsevier Ltd, 157–169. ����: 0022-1694. https://www.cabdirect.org/ cabdirect/abstract/20133322462 (2021) (2013) (cit. on p. 34). 50. Hu�man, G. J. et al. The TRMM Multisatellite Precipitation Analysis (TMPA): QuasiGlobal, Multiyear, Combined-Sensor Precipitation Estimates at Fine Scales. Journal of Hydrometeorology 8. Publisher: American Meteorological Society Section: Journal of Hydrometeorology, 38–55. ����: 1525-7541, 1525-755X. https://journals.ametsoc.org/view/ journals/hydr/8/1/jhm560_1.xml (2021) (Feb. 1, 2007) (cit. on p. 34). 127 51. Yatagai, A. et al. A 44-Year Daily Gridded Precipitation Dataset for Asia Based on a Dense Network of Rain Gauges. Sola 5, 137–140 (2009) (cit. on p. 35). 52. Yang, X. et al. The Optimal Multimodel Ensemble of Bias-Corrected CMIP5 Climate Models over China. Journal of Hydrometeorology 21. Publisher: American Meteorological Society Section: Journal of Hydrometeorology, 845–863. ����: 1525-7541, 1525-755X. https: //journals.ametsoc.org/view/journals/hydr/21/4/jhm-d-19-0141.1.xml (2021) (Apr. 1, 2020) (cit. on pp. 35,37,121). 53. Zhenyu, H. a. N. & Tianjun, Z. Assessing the Quality of APHRODITE High-Resolution Daily Precipitation Dataset over Contiguous China. 59276c1479d15b66 36. Publisher: 59276c1479d15b66, 361–373. ����: 1006-9895. http:/ / www. iapjournals.ac . cn/dqkx/ en/article/doi/10.3878/j.issn.1006-9895.2011.11043 (2021) (2012) (cit. on p. 35). 54. Taylor, K. E., Stou�er, R. J. & Meehl, G. A. An Overview of CMIP5 and the Experiment Design. Bulletin of the American Meteorological Society 93. Publisher: American Meteorological Society Section: Bulletin of the American Meteorological Society, 485–498. https://journals.ametsoc.org/view/journals/bams/93/4/bams-d-11-00094.1.xml (2021) (Apr. 1, 2012) (cit. on p. 36). 55. Moss, R. H., Nakicenovic, N. & O’Neill, B. C. Towards New Scenarios for Analysis of Emissions, Climate Change, Impacts, and Response Strategies 132 pp. ����: 978-92-9169-125-8. http: //www.ipcc.ch/pdf/supporting-material/expert-meeting-report-scenarios.pdf (2021) (IPCC, Geneva, 2008) (cit. on p. 36). 56. Riahi, K. et al. RCP 8.5—A scenario of comparatively high greenhouse gas emissions. Climatic Change 109, 33. ����: 1573-1480. https://doi.org/10.1007/s10584-011-0149-y (2021) (Aug. 13, 2011) (cit. on p. 37). 57. Giorgi, F. et al. The Regional Climate Change Hyper-Matrix Framework. Eos, Transactions American Geophysical Union 89, 445–446. ����: 2324-9250. https://agupubs.onlinelibrary. wiley.com/doi/abs/10.1029/2008EO450001 (2021) (2008) (cit. on p. 37). 58. Giorgi, F. & Coppola, E. Does the model regional bias a�ect the projected regional climate change? An analysis of global model projections. Climatic Change 100, 787–795. ����: 1573-1480. https://doi.org/10.1007/s10584-010-9864-z (2021) (June 1, 2010) (cit. on pp. 37, 121). 59. Christensen, J. H., Kjellström, E., Giorgi, F., Lenderink, G. & Rummukainen, M. Weight assignment in regional climate models. Climate Research 44, 179–194. ����: 0936-577X, 1616-1572. https://www.int-res.com/abstracts/cr/v44/n2-3/p179-194/ (2021) (Dec. 9, 2010) (cit. on pp. 37,121). 60. Olmos Giménez, P., García Galiano, S. G. & Giraldo-Osorio, J. D. Identifying a robust method to build RCMs ensemble as climate forcing for hydrological impact models. Atmospheric Research 174-175, 31–40. ����: 0169-8095. https://www.sciencedirect.com/science/ article/pii/S0169809516300035 (2021) (June 15, 2016) (cit. on pp. 37,121). 61. Giorgi, F. et al. RegCM4: model description and preliminary tests over multiple CORDEX domains. Climate Research 52, 7–29. ����: 0936-577X, 1616-1572. https : / / www. int - res.com/abstracts/cr/v52/p7-29/ (2021) (Mar. 22, 2012) (cit. on p. 38). 128 62. Jones, C. D. et al. The HadGEM2-ES implementation of CMIP5 centennial simulations. Geoscienti�c Model Development 4. Publisher: Copernicus GmbH, 543–570. ����: 1991959X. https://gmd.copernicus.org/articles/4/543/2011/gmd-4-543-2011.html (2021) (July 1, 2011) (cit. on p. 38). 63. Ilyina, T. et al. Global ocean biogeochemistry model HAMOCC: Model architecture and performance as component of the MPI-Earth system model in di�erent CMIP5 experimental realizations. Journal of Advances in Modeling Earth Systems 5, 287–315. ����: 1942-2466. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2012MS000178 (2021) (2013) (cit. on p. 39). 64. EasyBal 2021. https://h2ogeo.upc.edu/en/investigation-hydrogeology/software/147-easybal-en (2021) (cit. on pp. 41,55). 65. Guo, Z. EasyBal Water Balance Calculations for 1995-2005 over Wanzhou County, China. 2021 (cit. on pp. 41,57–59,117). 66. Fischer, M., Rust, H. W. & Ulbrich, U. Seasonal Cycle in German Daily Precipitation Extremes. Meteorologische Zeitschrift. Publisher: Schweizerbart’sche Verlagsbuchhandlung, 3–13. ����:,https://www.schweizerbart.de/papers/metz/detail/27/88248/Seasonal_Cycle_ in_German_Daily_Precipitation_Extre?af=crossref (2021) (Jan. 29, 2018) (cit. on pp. 41, 118). 67. Rust, H. W., Maraun, D. & Osborn, T. J. Modelling seasonality in extreme precipitation. The European Physical Journal Special Topics 174, 99–111. ����: 1951-6401. https://doi.org/ 10.1140/epjst/e2009-01093-7 (2021) (July 1, 2009) (cit. on pp. 41,118). 68. Delignette-Muller, M. L. & Dutang, C. �tdistrplus: An R Package for Fitting Distributions. Journal of Statistical Software 64. Number: 1, 1–34. ����: 1548-7660. https://www.jstatsoft. org/index.php/jss/article/view/v064i04 (2021) (Mar. 20, 2015) (cit. on p. 42). 69. Laio, F. Cramer–von Mises and Anderson-Darling goodness of �t tests for extreme value distributions with unknown parameters. Water Resources Research 40. ����: 1944-7973. https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2004WR003204 (2021) (2004) (cit. on pp. 43,44). 70. Marsaglia, G. & Marsaglia, J. Evaluating the Anderson-Darling Distribution. Journal of Statistical Software 9. Number: 1, 1–5. ����: 1548-7660. https://www.jstatsoft.org/index. php/jss/article/view/v009i02 (2021) (Feb. 25, 2004) (cit. on p. 43). 71. Xu, L. & Wang, A. Application of the Bias Correction and Spatial Downscaling Algorithm on the Temperature Extremes From CMIP5 Multimodel Ensembles in China. Earth and Space Science 6, 2508–2524. ����: 2333-5084. https://agupubs.onlinelibrary.wiley.com/doi/ abs/10.1029/2019EA000995 (2021) (2019) (cit. on pp. 46,107,120,121). 72. Cannon, A. J., Sobie, S. R. & Murdock, T. Q. Bias Correction of GCM Precipitation by Quantile Mapping: How Well Do Methods Preserve Changes in Quantiles and Extremes? Journal of Climate 28. Publisher: American Meteorological Society Section: Journal of Climate, 6938–6959. ����: 0894-8755, 1520-0442. https://journals.ametsoc.org/view/ journals/clim/28/17/jcli-d-14-00754.1.xml (2021) (Sept. 1, 2015) (cit. on p. 46). 129 73. Maraun, D. Bias Correcting Climate Change Simulations - a Critical Review. Current Climate Change Reports 2, 211–220. ����: 2198-6061. https://doi.org/10.1007/s40641-0160050-x (2021) (Dec. 1, 2016) (cit. on p. 47). 74. Kim, S., Shin, J. - Y., Ahn, H. & Heo, J. - H. Selecting Climate Models to Determine Future Extreme Rainfall Quantiles. Journal of the Korean Society of Hazard Mitigation 19. Publisher: Korean Society of Hazard Mitigation, 55–69. ����: 1738-2424, 2287-6723. http://www.jkosham.or.kr/journal/view.php?doi=10.9798/KOSHAM.2019.19.1.55 (2021) (Feb. 28, 2019) (cit. on p. 47). 75. Durman, C. F., Gregory, J. M., Hassell, D. C., Jones, R. G. & Murphy, J. M. A comparison of extreme European daily precipitation simulated by a global and a regional climate model for present and future climates. Quarterly Journal of the Royal Meteorological Society 127, 1005–1015. ����: 1477-870X. https://rmets.onlinelibrary.wiley.com/doi/abs/10.1002/qj. 49712757316 (2021) (2001) (cit. on p. 47). 76. Li, J., Wasko, C., Johnson, F., Evans, J. P. & Sharma, A. Can Regional Climate Modeling Capture the Observed Changes in Spatial Organization of Extreme Storms at Higher Temperatures? Geophysical Research Letters 45, 4475–4484. ����: 1944-8007. https : / / agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2018GL077716 (2021) (2018) (cit. on p. 48). 77. Kim, S., Joo, K., Kim, H., Shin, J. - Y. & Heo, J. - H. Regional quantile delta mapping method using regional frequency analysis for regional climate model precipitation. Journal of Hydrology 596, 125685. ����: 0022-1694. https://www.sciencedirect.com/science/article/ pii/S002216942031146X (2021) (May 1, 2021) (cit. on pp. 48,49). 78. Nguyen, H., Mehrotra, R. & Sharma, A. Can the variability in precipitation simulations across GCMs be reduced through sensible bias correction? Climate Dynamics 49, 3257– 3275. ����: 1432-0894. https://doi.org/10.1007/s00382-016-3510-z (2021) (Nov. 1, 2017) (cit. on p. 49). 79. Maraun, D. & Widmann, M. Cross-validation of bias-corrected climate simulations is misleading. Hydrology and Earth System Sciences 22. Publisher: Copernicus GmbH, 4867– 4873. ����: 1027-5606. https://hess.copernicus.org/articles/22/4867/2018/ (2021) (Sept. 18, 2018) (cit. on p. 50). 80. Koutsoyiannis, D. Statistics of extremes and estimation of extreme rainfall: I. Theoretical investigation. Hydrological Sciences Journal 49, 3. ����: 0262-6667, 2150-3435. https://www. tandfonline.com/doi/full/10.1623/hysj.49.4.575.54430 (2021) (Aug. 2004) (cit. on pp. 95, 118). 81. Sandvik, M. I., Sorteberg, A. & Rasmussen, R. Sensitivity of historical orographically enhanced extreme precipitation events to idealized temperature perturbations. Climate Dynamics 50, 143–157. ����: 1432-0894. https://doi.org/10.1007/s00382-017-3593-1 (2021) (Jan. 1, 2018) (cit. on p. 108). 82. Napoli, A., Crespi, A., Ragone, F., Maugeri, M. & Pasquero, C. Variability of orographic enhancement of precipitation in the Alpine region. Scienti�c Reports 9, 13352. ����: 20452322. https://www.nature.com/articles/s41598-019-49974-5 (2021) (Sept. 16, 2019) (cit. on p. 108). 130 83. Miura, T. & Nagai, S. Landslide Detection with Himawari-8 Geostationary Satellite Data: A Case Study of a Torrential Rain Event in Kyushu, Japan. Remote Sensing 12. Number: 11 Publisher: Multidisciplinary Digital Publishing Institute, 1734. https://www.mdpi.com/ 2072-4292/12/11/1734 (2021) (Jan. 2020) (cit. on p. 117). 84. Tseng, C. - W. et al. Application of High-Resolution Radar Rain Data to the Predictive Analysis of Landslide Susceptibility under Climate Change in the Laonong Watershed, Taiwan. Remote Sensing 12. Number: 23 Publisher: Multidisciplinary Digital Publishing Institute, 3855. https://www.mdpi.com/2072-4292/12/23/3855 (2021) (Jan. 2020) (cit. on p. 117). 85. Kumar, P. et al. Towards an operationalisation of nature-based solutions for natural hazards. Science of The Total Environment 731, 138855. ����: 0048-9697. https://www.sciencedirect. com/science/article/pii/S004896972032372X (2021) (Aug. 20, 2020) (cit. on p. 119). 86. Wood, A. W., Leung, L. R., Sridhar, V. & Lettenmaier, D. P. Hydrologic Implications of Dynamical and Statistical Approaches to Downscaling Climate Model Outputs. Climatic Change 62, 189–216. ����: 1573-1480. https://doi.org/10.1023/B:CLIM.0000013685.99609.9e (2021) (Jan. 1, 2004) (cit. on p. 120). 131