Attributing a deadly landslide disaster in Southeastern Brazil to human-induced climate change – Supplementary Material Maria Lucia Ferreira Barbosa1,2, Rafaela Quintella Veiga3, Renata Pacheco Quevedo4, Débora Joana Dutra1, Ana Carolina Moreira Pessôa5, Thaís Pereira de Medeiros1, Chantelle Burton6, Yuexiao Liu7, Nubia Beray Armond3, Rafael C. de Abreu7, Sihan Li8, Fraser C. Lott6, Cassiano Antonio Bortolozo9, Sarah Sparrow7, Liana Oighenstein Anderson1 1Earth Observation and Geoinformatics Division, National Institute for Space Research-INPE, Tropical Ecosystems and Environmental Sciences Laboratory (TREES), São José dos Campos, SP, 12227-010, Brazil. 2UK Centre for Ecology & Hydrology, Wallingford OX10 8BB, U.K 3 Department of Geography, Indiana University (IU), Blooomington, Indiana, United States. 4 ENGAGE Research Group, Department of Geography and Regional Research, University of Vienna, Vienna 1010, Austria. 5Instituto de Pesquisa Ambiental da Amazônia - IPAM, Brasília, 70863-520, Brazil. 6 Met Office Hadley Centre, Met Office, UK. 7 Engineering Science, University of Oxford, Oxford, UK. Mini TESA, Holywell House, Osney Mead, OX2 0ES, Oxford e-Research Centre. Tel: +441865 610682 8 University of Sheffield, UK 9 Universidade de São Paulo (USP), Instituto de Astronomia, Geofísica e Ciências Atmosféricas (IAG), Departamento de Geofísica, Rua do Matão, 1226, Butantã, 05508-090 São Paulo, Brazil 10 Divisão de Observação da Terra e Geoinformática (DIOTG), Coordenação Geral de Ciências da Terra (CGCT), Instituto Nacional de Pesquisas Espaciais (INPE) Corresponding author: Sarah Sparrow
[email protected] 1. Land-use and land-cover change in landslide-prone areas of Petrópolis, Brazil (1985–2024) Supplementary Figure 1 synthesizes nearly four decades of land-use and land-cover (LULC) dynamics within mapped landslide points in the municipality of Petrópolis, Rio de Janeiro, Brazil. Panel (a) presents a Sankey diagram that traces the temporal transitions of each landslide location among the ten MapBiomas classes from 1985 to 2012 and 2022. The width of the connecting bands is proportional to the number of points undergoing each transition, revealing a pronounced flow from Forest Formation in 1985 to Urban cover in subsequent years, which underscores the expansion of built-up areas over formerly forested slopes. Panel (b) displays the absolute frequency of landslide points per LULC class in the three reference years, showing a steady increase in urban cover accompanied by a decline in forested categories. Panel (c) depicts the relative landslide risk, expressed as the proportion of total mapped points occurring in each class, and demonstrates a marked rise in risk within urban areas concurrent with a reduction across natural vegetation types. Collectively, these results highlight the progressive urbanization of steep terrain and its close association with the growing concentration of landslide events in Petrópolis over the 1985–2022 period.
Figure S1 – Land-use and land-cover change in landslide-prone areas of Petrópolis, Brazil (1985– 2022). (a) Sankey diagram showing how the land-cover class of each mapped landslide point changed over time. Flows connect the class occupied in 1985 (left) to that in 2012 (centre) and 2022 (right). The width of each flow is proportional to the number of landslide points following that transition. (b) Bar chart of the total number of landslide points within each class for 1985, 2012 and 2022. (c) Relative landslide risk for each class, expressed as the proportion of total points occurring in that class at each date. 2. Project Overview This repository documents the full analytical pipeline used to quantify and model landuse and land-cover (LULC) changes from 1985 to 2024 in Petrópolis, Rio de Janeiro,
Brazil, with a focus on areas prone to landslides and hilltop Permanent Preservation Areas (APPs). The workflow retrieves annual LULC data from MapBiomas Collection 10, reclassifies it into ten consolidated categories, calculates annual class areas for the entire municipality and for mapped landslide points, and applies Bayesian time-series modelling to detect long-term trends. All outputs—including CSV tables, graphics, and posterior distributions—are reproducible with the code provided. 3. Repository Contents A single Python script (Colab/Jupyter compatible) implementing every stage of the analysis: • Google Earth Engine (GEE) Access Authenticates the user and initializes the Earth Engine API with the project ee- • Data Acquisition o Municipal boundary: custom GEE asset brazil/municipios. o Landslide inventory: custom GEE asset landslide_petropolis. o Elevation: USGS SRTM 30 m DEM for slope and hilltop delineation. o Land-cover: MapBiomas Brazil Collection 10 (annual 30 m maps, 1985– 2023/24). • Pre-processing o Reclassifies the 60+ MapBiomas classes into 10 analytical categories Forest Formation, Forest Plantation, Herbaceous/Shrubby Vegetation, Other Forest Types, Pasture, Temporary Crops, Perennial Crops, Urban, Water, Other. o Creates binary class bands and masks hilltop APPs (≥ 100 m relative height and ≥ 25° slope). • Area Computation o Annual total area (km²) of each class for: 1. the entire municipality; 2. every mapped landslide point (using reduceRegions). o Exports annual CSV files to Google Drive. • Aggregation & Transition Analysis o Merges yearly CSVs into a single long table (landuse_aggregated.csv). o Produces per-landslide-type summaries. o Builds transition matrices and Sankey diagrams (1985→2012→2022). • Time-series and Statistical Modelling o Converts class areas to percentages of total landscape. o Fits a Bayesian hierarchical linear model (PyMC) for each class: o Outputs posterior means, 90 % highest-density intervals (HDI), slope probabilities (increase/decrease), and posterior-predicted trends. • Visualization Generates all figures used in the article and supplements: Sankey diagrams, bar plots of class frequencies, relative-risk plots, forest plots of slope estimates, annual gain/loss charts, HDI trend curves, and heatmaps of annual percentage change. 4. Dataset Description
The analysis produces a structured, openly reusable dataset stored in Google Drive folders created by the script: • Annual Class Areas (landuse_sample/landuse_area_YYYY.csv) One file per year (1985–2024). Columns: Forest Formation, Forest Plantation, …, Other (areas in km²). • Landslide-Point Class Areas (landuse_petropolis/landuse_area_points_YYYY.csv) Class areas (m²) intersecting each landslide point, with attributes: FID (unique landslide ID), Tipo (landslide type), year, and class_1–class_10. • Aggregated Tables o landuse_aggregated.csv – total class areas by landslide type and year. o by_type/landuse_aggregated_<Tipo>.csv – per-type annual summaries. • Model Outputs o Posterior samples (PyMC trace) for intercepts and slopes. o Summary metrics: RMSE, MAE, R², Pseudo-R² (per class and global). o Derived time-series of posterior mean percentages and HDI intervals. Each CSV is in UTF-8 encoding with comma separation, ready for import into R, Python, or GIS software. 5. Requirements • Python ≥ 3.9 • Key libraries: earthengine-api, geemap, pandas, numpy, matplotlib, seaborn, plotly, pymc (v4+), arviz, scikit-learn, ipyleaflet. • Google Earth Engine account with access to the specified assets. • Optional: Google Colab with Google Drive for automatic file export. 6. Execution Guide • Clone or upload the repository to a Google Colab notebook or local Jupyter environment. • Install dependencies (Colab example): • pip install earthengine-api geemap pymc arviz plotly seaborn • Authenticate and initialize Earth Engine when prompted (ee.Authenticate()). • Run codigo_petropolis.py. • CSV tables will be written to MyDrive/landuse_sample and MyDrive/landuse_petropolis. • All figures and model results are generated automatically.
