Repositorio Institucional de Documentos

2021 54 Adrián Jiménez Ruano Characterizing pyroregions in mainland Spain from spatialtemporal patterns of fire regime and their underlying drivers Departamento Director/es Geografía y Ordenación del Territorio Rodrigues Mimbrero, Marcos De la Riva Fernández, Juan © Universidad de Zaragoza Servicio de Publicaciones ISSN 2254-7606 Adrián Jiménez Ruano CHARACTERIZING PYROREGIONS IN MAINLAND SPAIN FROM SPATIAL-TEMPORAL PATTERNS OF FIRE REGIME AND THEIR UNDERLYING DRIVERS Director/es Geografía y Ordenación del Territorio Rodrigues Mimbrero, Marcos De la Riva Fernández, Juan Tesis Doctoral Autor 2019 Repositorio de la Universidad de Zaragoza – Zaguan http://zaguan.unizar.es UNIVERSIDAD DE ZARAGOZA Tesis Doctoral Characterizing pyroregions in mainland Spain from spatial-temporal patterns of fire regime and their underlying drivers Autor Adrián Jiménez Ruano Director/es de la Riva Fernández, Juan Rodrigues Mimbrero, Marcos Geografía y Ordenación del Territorio 2019 El autor de esta tesis doctoral disfrutó para su desarrollo, así como para la estancia en un centro de investigación extranjero, de la financiación del Programa de ayudas FPU del Ministerio de Educación, Cultura y Deporte. Referencia de la ayuda FPU 13/06618. Desea hacer constar, por tanto, su agradecimiento. Jiménez-Ruano A, Rodrigues M, Jolly W M, de la Riva Fernández J. (2018). The role of drought and magnitude in the temporal evolution of fire occurrence and burned area size in mainland Spain. Geophysical Research Abstracts (Poster contribution). Vol. 20, EGU2018-13520, Vienna, Austria. In this work, the PdD student, Adrián Jiménez Ruano, is responsible for most of the work, having carried out the statistical and spatial analysis, and being the main responsible for elaborating the poster. Dr. Marcos Rodrigues and Dr. Juan de la Riva have collaborated very closely in the conception of the methodology and in the review of the poster. Dr. Matt Jolly has also contributed to the methodology design as well as to the review of the results. La presente tesis doctoral se ha elaborado siguiendo la modalidad de compendio de publicaciones. El doctorando, Adrián Jiménez Ruano, figura como primer autor y responsable de casi todos los artículos. Seguidamente se detallan los trabajos que constituyen el cuerpo de la tesis, su factor de impacto y el detalle de las tareas realizadas por cada uno de los autores en cada uno de ellos: Jiménez-Ruano A, Rodrigues M, de la Riva J (2017) Understanding wildfires in mainland Spain. A comprehensive analysis of fire regime features in a climate-human context. Applied Geography 89:100-111. https://doi.org/10.1016/j.apgeog.2017.10.007 Factor de impacto JCR: 3,117 (1er cuartil, “Geography”). En este trabajo el doctorando, Adrián Jiménez Ruano, es responsable de la mayor parte del trabajo, desarrollando gran parte del análisis estadístico y espacial, siendo el responsable último de la redacción de los contenidos. El Dr. Marcos Rodrigues Mimbrero ha colaborado en el desarrollo de algunas tareas y también ayudó en el proceso de escritura. Tanto el Dr. Juan de la Riva and Marcos Rodrigues Mimbrero figuran como coautores en calidad de directores de tesis, siendo responsables del tema general de la investigación y habiendo colaborado en la revisión de los resultados. Jiménez-Ruano A, Rodrigues M, de la Riva J (2017) Exploring spatial–temporal dynamics of fire regime features in mainland Spain. Natural Hazards and Earth System Sciences 17:1697-1711. https://doi.org/10.5194/nhess-17-1697-2017 Factor de impacto JCR: 2,281 (2º cuartil, “Geosciences, Multidisciplinary”). En este trabajo el doctorando, Adrián Jiménez Ruano, es el responsable de la mayor parte del trabajo, desarrollando gran parte del análisis estadístico y espacial, siendo el responsable último de la redacción de los contenidos. El Dr. Marcos Rodrigues Mimbrero y el Dr. Juan de la Riva también han colaborado muy estrechamente en la revisión de la metodología y los resultados. Jiménez-Ruano A, Rodrigues M, Jolly W.M, de la Riva J (2019) The role of short-term weather conditions in temporal dynamics of fire regime features in mainland Spain. Journal of Environmental Management 241:575-586. https://doi.org/10.1016/j.jenvman.2018.09.107 Factor de Impacto JCR: 4,865 (1er cuartil, “Environmental Sciences”). En este trabajo el doctorando, Adrián Jiménez Ruano, es el responsable de la mayor parte del trabajo, desarrollando gran parte del análisis estadístico y espacial, siendo el responsable último de la redacción de los contenidos. El Dr. Matt Jolly figura como coautor por su colaboración en la construcción de parte de la metodología y revisión de los resultados preliminares siendo además el responsable de la estancia en la que se desarrolló la investigación. El Dr. Marcos Rodrigues Mimbrero y el Dr. Juan de la Riva también han colaborado muy estrechamente en la revisión de los resultados. Rodrigues M, Jiménez-Ruano A, de la Riva J. (2016) Analysis of recent spatial–temporal evolution of human driving factors of wildfires in Spain. Natural Hazards 84(3):2049-2070. https://doi.org/10.1007/s11069-016-2533-4 Factor de impacto JCR: 1,833 (2º cuartil, “Geosciences, Multidisciplinary”). En este trabajo, el Dr. Marcos Rodrigues es responsable de la mayor parte del trabajo, habiendo realizado el análisis estadístico y espacial, y siendo el principal responsable de la redacción de todos los contenidos. El doctorando Adrián Jiménez Ruano ha colaborado en la ejecución de parte de la metodología. El Dr. Juan de la Riva es el responsable del tema general de la investigación y ha colaborado en la revisión de los resultados. Rodrigues M, Jiménez-Ruano A, Peña-Angulo D, de la Riva J. (2018) A comprehensive spatial-temporal analysis of driving factors of human-caused wildfires in Spain using Geographically Weighted Logistic Regression. Journal of Environmental Management 225: 177-192. https://doi.org/10.1016/J.JENVMAN.2018.07.098 Factor de impacto JCR: 4,865 (1º cuartil, “Environmental Sciences”). En este trabajo, el Dr. Marcos Rodrigues es responsable de la mayor parte del trabajo, habiendo realizado el análisis estadístico y espacial, y siendo el principal responsable de la redacción de todos los contenidos. El doctorando, Adrián Jiménez Ruano, ha colaborado en la realización de parte de la metodología y parte de la cartografía del trabajo final. La Dr. Peña-Angulo ha proporcionado los datos climáticos empleados en la investigación. El Dr. Juan de la Riva es el responsable del tema general de la investigación y ha colaborado en la revisión de los resultados. Además, se han incluido otros trabajos en el cuerpo de la investigación, pero que aún no han sido publicados: Rodrigues M, Jiménez-Ruano A, de la Riva J (En revisión). Fire regime dynamics in mainland Spain. Part 1: drivers of change. Science of the Total Environment. Factor de impacto JCR: 5,589 (1º cuartil, “Environmental Sciences”). En este trabajo el Dr. Marcos Rodrigues Mimbrero, es el responsable de la mayor parte del trabajo, desarrollando gran parte del análisis estadístico y especial, siendo el responsable último de la redacción de los contenidos. El doctorando, Adrián Jiménez Ruano, ha colaborado muy estrechamente en la confección del proceso metodológico y en obtener parte de los resultados, y el Dr. Juan de la Riva también ha participado activamente en la revisión de los resultados. Jiménez-Ruano A, de la Riva J, Rodrigues M (En revisión). Fire regime dynamics in mainland Spain. Part 2: a near-future prospective of fire activity. Science of the Total Environment. Factor de impacto JCR: 5,589 (1º cuartil, “Environmental Sciences”) En este trabajo el doctorando, Adrián Jiménez Ruano, es el responsable de la mayor parte del trabajo, desarrollando gran parte del análisis estadístico y espacial, siendo el responsable último la redacción de los contenidos. El Dr. Marcos Rodrigues Mimbrero ha colaborado muy estrechamente en la confección del proceso metodológico final y revisión del texto, y el Dr. Juan de la Riva también ha participado estrechamente en la revisión de los resultados. Por último, se han incluido varias contribuciones de congresos en dos apéndices específicos: Apéndice D Jiménez-Ruano A, Rodrigues M, de la Riva Fernández J. 2018. Identifying pyroregions by means of Self Organizing Maps and hierarchical clustering algorithms in mainland Spain, in: Viegas, D.X. (Ed.), Advances in Forest Fire Research (VIII International Conference on Forest Fire Research). Imprensa da Universidade de Coimbra, Coimbra, pp. 495–505. https://doi.org/https://doi.org/10.14195/978-989-26-16-506_54 En este trabajo, el doctorando Adrián Jiménez Ruano es el responsable de la mayor parte del trabajo, habiendo realizado el análisis estadístico y espacial, y siendo el principal responsable de la redacción de todos los contenidos. El Dr. Marcos Rodrigues ha colaborado muy estrechamente en la preparación del proceso metodológico y en la revisión de la redacción. El Dr. Juan de la Riva también ha participado estrechamente en la revisión de los resultados. Apéndice E Jiménez-Ruano A, Rodrigues M, de la Riva Fernández J. (2017). An analysis of wildfire frequency and burned area relationships with human pressure and climate gradients in the context of fire regime. Geophysical Research Abstracts (Poster contribution). Vol. 19 EGU2017-15084, Vienna, Austria. En este trabajo, el doctorando, Adrián Jiménez Ruano, es el responsable de la mayor parte del trabajo, habiendo realizado el análisis estadístico y espacial, y siendo el principal responsable de la elaboración del póster. El Dr. Marcos Rodrigues ha colaborado muy estrechamente en la concepción de la metodología y en la revisión del diseño del póster. El Dr. Juan de la Riva también ha participado en la revisión de los resultados. Jiménez-Ruano A, Rodrigues M, de la Riva Fernández J. (2017). Assessing the influence of small fires on trends in fire regime features at mainland Spain. Geophysical Research Abstracts (Poster contribution). Vol. 19, EGU2017-15755, Vienna, Austria. En este trabajo, el doctorando, Adrián Jiménez Ruano, es el responsable de la mayor parte del trabajo, habiendo realizado el análisis estadístico y espacial, y siendo el principal responsable de la elaboración del póster. El Dr. Marcos Rodrigues ha colaborado muy estrechamente en la concepción de la metodología y en la revisión del diseño del póster. El Dr. Juan de la Riva también ha participado en la revisión de los resultados. Jiménez-Ruano A, Rodrigues M, Jolly W M, de la Riva Fernández J. (2018). Assessing the influence of fire weather danger indexes on fire frequency and burned area in mainland Spain. Geophysical Research Abstracts (Oral presentation). Vol. 20, EGU2018-13196, Vienna, Austria. En este trabajo, el doctorando, Adrián Jiménez Ruano, es el responsable de la mayor parte del trabajo, habiendo realizado el análisis estadístico y espacial, y siendo el principal responsable de la preparación de la presentación. El Dr. Marcos Rodrigues y el Dr. Juan de la Riva han colaborado muy estrechamente en la concepción de la metodología y en la revisión de la comunicación oral. El Dr. Matt Jolly también ha contribuido al cálculo de los índices meteorológicos de incendios, así como a la revisión de los resultados. Jiménez-Ruano A, Rodrigues M, Jolly W M, de la Riva Fernández J. (2018). The role of drought and magnitude in the temporal evolution of fire occurrence and burned area size in mainland Spain. Geophysical Research Abstracts (Poster contribution). Vol. 20, EGU2018-13520, Vienna, Austria. En este trabajo, el doctorando, Adrián Jiménez Ruano, es el responsable de la mayor parte del trabajo, habiendo realizado el análisis estadístico y espacial, y siendo el principal responsable de la elaboración del póster. El Dr. Marcos Rodrigues y el Dr. Juan de la Riva han colaborado muy estrechamente en la concepción de la metodología y en la revisión del póster. El Dr. Matt Jolly también ha contribuido al diseño de la metodología, así como a la revisión de los resultados. ABSTRACT Fire has always been an intrinsic feature in various ecosystems around the world. In environments heavily populated by humans, their actions have altered these natural fire regimes for others that are fundamentally anthropogenic in nature. In the context of Mediterranean Europe, the number of forest fires and their observed burnt area fell into a general decline during the late twentieth century, which led to a reduced incidence of fire in most Mediterranean ecosystems historically affected by recurrent fires. Therefore, the change in past fire regimes is evident, mainly due to human intervention instigating a very demanding policy of total exclusion of fire. However, the recent evolution of fire regimes presents a high spatial and temporal variability. On the other hand, future scenarios predict a growing impact of the human factor (more land abandonment, poor management of forests and adhering exclusively to suppression methods), which will result in increased fire activity due to a greater amount of available fuel. In addition, climatic conditions are expected to cause increasingly larger burned areas (higher temperatures, more frequent heat waves and droughts), which will undoubtedly have a negative effect on both ecosystems and future societies. All these factors make an adequate zoning of fire regimes necessary from a spatial-temporal perspective, which allows the relationship between the altered fire regime and associated socio-economic and environmental factors to be determined, as well as detecting temporal trends in regions with decreasing activity, or on the contrary, an increase in the incidence of fires. Therefore, finding these areas will lead to improved management and prevention of forest fires. This doctoral dissertation focuses on enriching knowledge for identifying and interpreting homogeneous regions of fire regimes. A wide range of methods of statistical analysis and spatial modeling are employed. The dissertation is structured according to the following objectives: Objective 1 focuses on analyzing the spatial-temporal distribution of the main features defining the fire regime during the recent period. Objective 2 aims to further describe the influence of meteorological danger on the evolution of fire activity. Objective 3 evaluates the change in the relative contribution of anthropogenic factors on forest fires. Objective 4 focuses on explaining the evolution and causes of changes or transitions in fire regimes during the recent (1974-2015) and future (2016-2036) periods. Finally, Objective 5 centers on the transfer of the zoning of fire regime typologies into an integral mapping of pyroregions The results indicate that fire regimes in mainland Spain have undergone several changes, mainly a considerable decrease in fire activity in most of the territory, although it still remains high in the north (especially in winter). The diverse machine-learning methods employed, especially Random Forest, have demonstrated their potential in terms of revealing the fire drivers behind fire regime evolution. Moreover, forecasting by the ARIMA model has confirmed the ongoing tendency towards a lower incidence of fire. All indications are that preventive measures should take greater prominence in areas with an abrupt decrease in wildfires, as they are significantly more prone to large ones in the short and medium term. RESUMEN El fuego ha coexistido de forma intrínseca en diversos ecosistemas a nivel global. En el caso de los ambientes más humanizados la acción del hombre ha alterado esos regímenes de incendio naturales por uno fundamentalmente de carácter antrópico. En el contexto de la Europa Mediterránea, el número de incendios forestales y su área quemada observados han experimentado un descenso general durante el final del siglo XX. Esto ha supuesto un declive de la incidencia del fuego en la mayoría de los ecosistemas mediterráneos históricamente afectados por incendios recurrentes. Por tanto, es evidente la alteración de los regímenes de incendio pasados, debido principalmente a la intervención humana con una política de exclusión total del fuego muy exigente. No obstante, la evolución reciente de los regímenes de incendio presenta una alta variabilidad espacial y temporal. Por otro lado, las perspectivas de futuro vaticinan un impacto creciente del factor humano (abandono del campo, gestión de los bosques y mantenimiento de la supresión excluyente), lo que consecuentemente derivará una mayor actividad de incendios debido a una mayor cantidad de combustible disponible. Asimismo, se prevén unas condiciones climáticas cada vez más propensas a generar incendios de gran superficie (mayores valores de temperatura, mayor frecuencia de olas de calor y sequías), lo que sin duda afectará negativamente tanto a los ecosistemas como las sociedades futuras. Todos estos factores hacen necesaria una adecuada zonificación de los regímenes de incendio desde una perspectiva espacio-temporal, la cual permita conocer la relación existente entre el régimen de incendios alterado y los factores socio-económicos y ambientales asociados. Así como detectar tendencias en el tiempo en regiones que experimenten un descenso de la actividad, o, por el contrario, incremento de la incidencia de incendios. Por tanto, conociendo estas zonas se podrá mejorar la gestión y prevención contra incendios forestales. Esta tesis doctoral se enfoca en enriquecer el conocimiento sobre la identificación e interpretación de regiones homogéneas de regímenes de incendio. Para ello se recurre a un amplio abanico de métodos de análisis estadísticos y de modelado espacial. La tesis se estructura de acuerdo a los siguientes objetivos: el objetivo 1 se centra en analizar la distribución espacio-temporal de las principales métricas que definen el régimen de incendio durante el periodo reciente. El objetivo 2 pretende profundizar en la influencia del riesgo meteorológico en la evolución de la actividad de los incendios. El objetivo 3 evalúa el cambio de la contribución relativa de los factores antropogénicos en los incendios forestales. El objetivo 4 se enfoca en explicar la evolución y causas de los cambios o transiciones de los regímenes de incendios durante el periodo reciente (1974-2015) y futuro (2016-2036). Finalmente, el objetivo 5 pone la atención en la traslación de la zonificación de tipologías de regímenes de incendios hacia una cartografía integral de piroregiones. Los resultados indican que los regímenes de incendio en la España peninsular han experimentado diversos cambios, principalmente una disminución considerable de la actividad de incendios en la mayor parte del territorio, aunque todavía persiste una alta actividad en el extremo norte (especialmente en invierno). Los diversos métodos de aprendizaje automático empleados, especialmente Random Forest, han demostrado su potencial en términos de revelar los factores que impulsan la evolución del régimen de incendios. Además, la proyección ARIMA ha confirmado la tendencia actual hacia una menor incidencia de incendios. Todo apunta a que las medidas preventivas deben tomar más protagonismo en áreas con un abrupto descenso de la ocurrencia, ya que son significativamente más propensas a grandes incendios a corto y medio plazo. TABLE OF CONTENTS CHAPTER 1: INTRODUCTION ........................................................................................................................ 1 1.1. The wildfire phenomenon ............................................................................................................................... 3 1.2. The concept of fire regime: definitions and components ......................................................................... 4 1.3. Methodological approaches in fire regime modelling ................................................................................ 5 1.4. Fire regime vs pyroregion ................................................................................................................................ 7 CHAPTER 2: OBJECTIVES AND RESEARCH DESIGN ...................................................................... 9 2.1. Research questions .......................................................................................................................................... 11 2.2. Research structure ........................................................................................................................................... 13 CHAPTER 3: STUDY AREA ............................................................................................................................... 15 CHAPTER 4: MATERIALS AND METHODS ........................................................................................... 21 4.1. Datasets and sources ....................................................................................................................................... 23 4.1.1. The Spanish fire database...................................................................................................................... 23 4.1.2. Fire data and fire features ...................................................................................................................... 24 4.1.3. Climate and weather ............................................................................................................................... 27 4.1.4. Anthropogenic drivers ........................................................................................................................... 29 4.2. Modelling techniques ...................................................................................................................................... 33 4.2.1. Descriptive and explorative .................................................................................................................. 33 4.2.2. Time series analysis ................................................................................................................................ 35 4.2.3. Classification and regression................................................................................................................. 39 CHAPTER 5: SPATIAL-TEMPORAL DISTRIBUTION OF FIRE REGIME FEATURES ..... 43 CHAPTER 6: THE INFLUENCE OF FIRE-WEATHER ON THE EVOLUTION OF FIRE ACTIVITY ........................................................................................................................ 73 CHAPTER 7: CHANGE IN ANTHROPOGENIC DRIVERS .............................................................. 87 CHAPTER 8: EVOLUTION AND CAUSES OF FIRE REGIME CHANGE ............................... 127 CHAPTER 9: TRANSLATING FIRE REGIME ZONING SCHEMES INTO PYROREGIONS ...................................................................................................................... 181 CHAPTER 10: CONCLUSIONS AND FUTURE RESEARCH .......................................................... 195 REFERENCES ....................................................................................................................................................... 207 APPENDIX A: SUPLEMENTARY MATERIAL OF FIRE REGIME FEATURES ................... 219 APPENDIX B: SUPLEMENTARY MATERIAL OF FIRE-WEATHER ........................................ 229 APPENDIX C: SUPLEMANTARY MATERIAL OF DRIVERS OF CHANGE ........................... 237 APPENDIX D: PRELIMINARY PYROREGIONS DELIMITATION ........................................... 251 APPENDIX E: CONFERENCES CONTRIBUTIONS ......................................................................... 265 Chapter 1: Introduction 6 specific-devoted simulation models, as change points (Mouillot et al., 2002) or power law (Malamud, 1998; Malamud et al., 2005; Perera and Cui, 2010). In Spain, several articles suggest that alterations in fire regimes have been driven by climate, land use changes and suppression policies (Moreno et al., 2014) as well as different propagation patterns in Catalonia (Duane et al., 2015). The majority of studies have used regression models in combination with simulated data from general climate models (GCM) (Boulanger et al. 2013; DaCamara et al. 2014; Kilpeläinen et al. 2010; Krawchuk et al. 2009; Pechony and Shindell 2010; Terrier et al. 2014; Westerling et al. 2011) based on IPPCC projections of future emission scenarios or Regional Climate Models (RCM). Most of these studies envisage an increasing burned area in regions such as Portugal (DaCamara et al., 2014), California (Westerling et al., 2011) and the Iberian Peninsula (Sousa et al., 2015). However, several authors point out different trends depending on the regions of the world (Krawchuk et al., 2009; Pechony and Shindell, 2010), including showing opposite tendencies with increasing frequency and a slight decline of burned area in the Northeast of Spain (Turco et al., 2014). In the fire-climate framework, many authors have analyzed the relationship between climate change and shifts in certain characteristics of fire regimes (fire frequency, surface area, seasonality, average fire range, maximum fire size, etc.) in many regions. For example, in the boreal forests of North America (Kasischke and Turetsky, 2006) they resort to historical records, the analysis of individual years by categories of ecozones and the start time of individual events. In Canada, the Fire Growth Model has been used to model the risk of lightning and human-induced ignitions (Nitschke and Innes, 2013). On the other hand, most of the studies have assumed future projections with similar environmental and anthropic conditions to the current ones (Boulanger et al., 2014, 2012), thus showing certain limitations in trend detection since they assume a “static” of non-climate conditions for the future. Therefore, the growing importance of estimating the present and future impact of climate change on fire regime has become a key issue in risk assessment and adaptation strategies, emerging as the cornerstone in national and international climate programs (Turco et al., 2014), such as the European project FUME (2010-2013). However, it is well-known that the democratic and massive use of future climate change scenarios implies a high degree of uncertainty. In other words, the most complicated issue is the validation of projected data, especially those by GGM or RCM models, as there is still no time series with which to correlate. This is why some authors leaned towards the “safest” alternatives, such as the auto-regression and moving average models (ARIMA). ARIMA models are known for their good performance in fields such as markets and the economy (Loi and Ng, 2018; Matyjaszek et al., 2019), as well as in the environmental framework: vegetation (REF) or climate change. In the context of forest fire, Preisler and Westerling (2007) employed ARIMA using temperature forecasting to assess fire danger in western USA, whereas, Boubeta et al. (2016) applied a simplified version of ARIMA (ARMA) without the integrated component in order to predict burnt area in Galicia. The main virtue of ARIMA models lies in the fact that they predict future trends and seasonality, with the historical time series of data as their only reference. As a result, the principle of parsimony is guaranteed in the model, since the minimum number of variables is used, being more easily reproducible and without creating an over adjustment. When discussing the use of spatial modeling methods and the prediction of forest fire characteristics in specific areas, a wide repertoire of methodological approaches and explanatory variables can be brought to bear. The scale of analysis (global, regional or local), the proposed objectives and the nature of the data used will condition the analysis framework. Among the most widely applied models to date, GLM and GAM (Generalized Linear Models and General Additive Models, respectively) stand out as flexible Chapter 1: Introduction 7 generalizations of linear regression, able to deal with non-normal distributions of the variable under study (fire activity) or the explanatory variables (climate, weather, topography, population, wildland urbanagricultural interfaces, road network, etc.). This modeling framework is adequate, since forest fire data usually depict non-linear response functions. In addition, Geographically Weighted Regression (GWR) is a more advanced alternative that has also been applied in the context of wildfires (Koutsias et al., 2010; Sá et al., 2011), whose main advantage is that it allows the calculation of local regression parameters, useful for analyzing the spatial behavior of each explanatory variable and determining their level of significance. The few works conducted in mainland Spain point to a certain degree of spatial variability (Martínez-Fernández et al., 2013), confirming that human driving factors vary over both space and time(Rodrigues et al., 2016) and are losing explanatory power in favor of climatic conditions (Rodrigues et al., 2018). Another important aspect when estimating the probability of the occurrence of wildfires is to analyze the characteristic of fuels and how they interact with climatic variables (precipitation, temperature, wind, relative humidity, etc.). The role of forest fuels not only largely determines the likelihood of ignition, but also the speed of propagation, and ultimately the severity. In this respect, numerous fire weather danger indices have been used to relate meteorological data to fire (Fire Weather Index: FWI, Standardized PrecipitationEvapotranspiration Index: SPEI, Palmer Drought Index: PDSI, among others). Some studies carried out in Portugal (Fernandes et al., 2014) stressed the positive relationship between fire and weather together with fuel hazard and the final burned area. In eastern-Spain, (Cardil et al., 2019) pointed out the importance of a multi-temporal perspective when studying the link between drought and burned area for different vegetation communities. In any case, that there is a strong influence from the flammability of the fuel and its spatial continuity on fire frequency and burned area has been demonstrated by fuel model classifications (Prometheus, NFFL, NFDRS, McArthur, FBP - see Arroyo, Pascual, and Manzanera (2008) for more details. 1.4. Fire regime vs pyroregion Because of the potential usefulness and interest in predicting how the behavior of fire regimes evolves, this PhD Thesis aims to optimize the identification and characterization of fire regimes in mainland Spain, beginning with the identification of the most relevant features of fire, and continuing by evaluating the direction and extent of regional trends both in space and time. Until now, most works focused on broadscale fire regime modeling based on large ecological and administrative units. In Canada, the next step was to outline homogeneous fire regime (HFR) zones without this traditional approach, since it does not capture the spatial heterogeneity of fire regimes and could lead to spatially inaccurate estimations of future fire activity (Boulanger et al., 2014). In Spain, only a few papers have defined fire regime units but by attributing a static image in their delimitation (Moreno and Chuvieco, 2013), i.e., not incorporating the non-stationary behavior of fire features. To overcome this limitation, our goal is to provide a projection of the possible future evolution of these homogeneous fire regime zones, since until now the few projects carried out in the Iberian Peninsula have focused on only forecasting selected components, such as the affected surface (Sousa et al., 2015). Generally, the spatial delimitation of fire regimes is based exclusively on the consideration of the main, defining features of fire. However, this zoning has to be incorporated into a more comprehensive spatial context that also integrates the driving factors (both climatic and human) most directly related to forest fires. In this respect, the first study that put forward this new concept was Fréjaville & Curt (2015), which added the concept of “pyroclimates” to the wildfire literature. They developed a new framework for analyzing regional changes in fire regimes from specific spatial-temporal patterns of forest fires and climate, Chapter 1: Introduction 8 defining it as a geographical entity displaying homogenous attributes with respect to fire regime, climate conditions (bioclimatic variables and fire danger indices) and the temporal trends of both. Therefore, we adopted part of the innovation of this concept in our term “pyroregions”, but in our case, we added human factors into the definition and not only the climate conditions. As result, we define pyroregion as “a geographical area sharing homogeneous fire regime features, climate-human conditions and the evolution of both”. To sum up, the main difference between “fire regime” and “pyroregion” lies fundamentally in the nature of their underlying factors. In the case of fire regime, it broadly refers to the average conditions in terms of fire features over time and space. The pyroregion transcends fire regime being a geographical entity that characterizes by uniform or homogeneous fire activity, but is also influenced by self-defining climatic and human conditions. 2 CHAPTER 2: OBJECTIVES AND RESEARCH DESIGN This chapter summarizes the objectives and structure of the thesis, connecting the former with their corresponding publications and appendices that compose the whole research. Chapter 2: Objectives and research design 11 The working hypothesis of this PhD Thesis is that mainland Spain presents different fire regimes defined by specific fire frequency, burned area, seasonality and cause, which are non-stationary over space and time, thus allowing modeling and envisaging their evolution. To understand the complexity of the phenomenon, we had to investigate the driving forces of fire regimes, which ultimately would lead to the definition of dynamic pyroregions, thus improving fire management, prevention and preparedness within a context of climate and socio-economic change. Therefore, the main objective of this research was to translate the variety of homogenous zones of fire regimes into pyroregions, providing insights into their possible evolution through the identification and characterization of their main components (frequency, size, seasonality, cause, etc.) and driving factors (climate, weather, human pressure, etc.). 2.1. Research questions In order to address the main objective stated before, five specific research questions (RQs) or objectives were formulated and addressed by studying several research papers. Table 1 shows the correlation between each specific objective and its corresponding publications. RQ 1: What is the spatial-temporal distribution of the main fire regime features and what is its relationships with climate-human factors? 1st Objective: Explore the spatial-temporal distribution of fire regime features and their relation with climate-human factors. RQ 2: What role does fire-weather danger play in the temporal evolution of fire regime features? 2nd Objective: Estimate the contribution of fire-weather danger on the observed evolution of fire activity. RQ 3: What have been the spatial-temporal changes in the influence of human factors on wildfires? 3rd Objective: Analysis spatial-temporal changes in the role of anthropogenic drivers on wildfires. RQ 4: What changes have been experienced by fire regimes and which factors are behind these dynamics? 4th Objective: Characterize the dynamics of recent-future fire regimes and know the drivers of their changes. RQ 5: How are pyroregions distributed in space on the basis of the observed evolution of fire regime typologies and drivers? 5th Objective: Translate the fire regime typologies scheme into pyroregions. Chapter 2: Objectives and research design 12 Table 1. Summary of the specific objectives and their corresponding publication or contribution. Objective Publication 1st Objective: Explore the spatial-temporal distribution of fire regime features and their relation with climate-human factors. CHAPTER 5 -Jiménez-Ruano A, Rodrigues M, de la Riva J (2017) Understanding wildfires in mainland Spain. A comprehensive analysis of fire regime features in a climate-human context. Applied Geography 89:100-111. https://doi.org/10.1016/j.apgeog.2017.10.007 -Jiménez-Ruano A , Rodrigues M, de la Riva J (2017) Exploring spatial–temporal dynamics of fire regime features in mainland Spain. Natural Hazards and Earth System Sciences 17:16971711. https://doi.org/10.5194/nhess-17-1697-2017 APENDIX A -Supplementary material from “Understanding wildfires in mainland Spain. A comprehensive analysis of fire regime features in a climate-human context”. APPENDIX E - Jiménez-Ruano A , Rodrigues M, de la Riva Fernández J. (2017). An analysis of wildfire frequency and burned area relationships with human pressure and climate gradients in the context of fire regime. Geophysical Research Abstracts (Poster contribution). Vol. 19 EGU2017-15084, Vienna, Austria. - Jiménez-Ruano A , Rodrigues M, de la Riva Fernández J. (2017). Assessing the influence of small fires on trends in fire regime features at mainland Spain. Geophysical Research Abstracts (Poster contribution). Vol. 19, EGU2017-15755, Vienna, Austria. 2nd Objective: Estimate the contribution of fireweather danger on the temporal evolution of fire activity. CHAPTER 6 -Jiménez-Ruano A , Rodrigues M, Jolly W.M, de la Riva J (In Press) The role of short-term weather conditions in temporal dynamics of fire regime features in mainland Spain. Journal of Environmental Management 17:1697-1711. https://doi.org/10.1016/j.jenvman.2018.09.107 APPENDIX B - Supplementary material from: “The role of short-term weather conditions in temporal dynamics of fire regime features in mainland Spain” APPENDIX E - Jiménez-Ruano A , Rodrigues M, Jolly W M, de la Riva Fernández J. (2018). Assessing the influence of fire weather danger indexes on fire frequency and burned area in mainland Spain. Geophysical Research Abstracts (Oral presentation). Vol. 20, EGU2018-13196, Vienna, Austria. - Jiménez-Ruano A , Rodrigues M, Jolly W M, de la Riva Fernández J. (2018). The role of drought and magnitude in the temporal evolution of fire occurrence and burned area size in mainland Spain. Geophysical Research Abstracts (Poster contribution). Vol. 20, EGU2018-13520, Vienna, Austria. 3rd Objective: Analysis of spatial-temporal changes in the role of anthropogenic drivers on wildfires. CHAPTER 7 -Rodrigues M, Jiménez-Ruano A , de la Riva J. (2016) Analysis of recent spatial–temporal evolution of human driving factors of wildfires in Spain. Natural Hazards 84(3):2049-2070. https://doi.org/10.1007/s11069-016-2533-4 -Rodrigues M, Jiménez-Ruano A , Peña-Angulo D, de la Riva J. (2018) A comprehensive spatial-temporal analysis of driving factors of human-caused wildfires in Spain using Geographically Chapter 2: Objectives and research design 13 2.2. Research structure The contents of the Thesis are organized as follows: Chapter 3 presents a description of the study area. Chapter 4 summarizes the data sources and methods employed in the research, complementing the information already published. Chapters 5 to 8 bring together the original version of accepted and published articles. Lastly, the last two chapters (Chapter 9 and 10) portray the final outline of pyroregions and summarize the main conclusions, respectively. Figure 1 summarizes the main databases and methodologies employed in the investigation according to the first four specific objectives. In addition, a complementary section provides further information, organized into five appendixes (A, B, C, D and E). The first three correspond to the supplementary material in three publications of the main body of this thesis, appendix A belongs to the paper entitled “Understanding wildfires in mainland Spain. A comprehensive analysis of fire regime features in a climate-human context”, appendix B is part of the article “The role of short-term weather conditions in temporal dynamics of fire regime features in mainland Spain” and appendix C corresponds to the manuscript under review “Fire regime dynamics in mainland Spain. Part 1: drivers of change”. Appendix D refers to Chapter 3 of the book “Advances in Forest Fire Research” edited by Domingos Xavier Viegas, as a result of the contribution in the “VIII International Conference on Forest Fire Research”, held in the city of Coimbra (Portugal) from 9 to 16 November 2018. The latter appendix includes several abstracts from different conference contributions held in EGU 2017 and EGU 2018. Weighted Logistic Regression. Journal of Environmental Management 225: 177-192. https://doi.org/10.1016/J.JENVMAN.2018.07.098 4th Objective: Characterize the dynamics of recent-future fire regimes and know the drivers of their changes. CHAPTER 8 -Rodrigues M, Jiménez-Ruano A , de la Riva J. (In press). Fire regime dynamics in mainland in Spain. Part 1: drivers of change. Science of the Total Environment. -Jiménez-Ruano A , de la Riva J, Rodrigues M. (In press). Fire regime dynamics in mainland Spain. Part 2: a near-future prospective of fire activity. Science of the Total Environment. APPENDIX C - Supplementary material from: “Fire regime dynamics in mainland in Spain. Part 1: drivers of change. Science of the Total Environment”. 5th Objective: Translate the fire regime typologies scheme into pyroregions. CHAPTER 9 -Jiménez-Ruano A , Rodrigues M, de la Riva J. (to be submitted) Mapping recent pyroregions on the basis of spatial-temporal patterns of fire regimes and environmental-human datasets in mainland Spain APPENDIX D - Jiménez-Ruano A, Rodrigues M, de la Riva J. (2018) Identifying pyroregions by means of Self Organizing Maps and hierarchical clustering algorithms in mainland Spain. in: Viegas, D.X. (Ed.), Advances in Forest Fire Research (VIII International Conference on Forest Fire Research). Imprensa da Universidade de Coimbra, Coimbra, pp. 495–505. https://doi.org/https://doi.org/10.14195/978-989-26-16-506_54 Chapter 2: Objectives and research design 14 Figure 1 Conceptual workflow of the thesis according to the first four specific objectives. 3 CHAPTER 3: STUDY AREA This chapter presents a description of the study area where the thesis has had its spatial framework. Chapter 4: Materials and methods 23 4.1. Datasets and sources 4.1.1. The Spanish fire database The General Statistics of Wildfires (Estadística General de Incendios Forestales: EGIF) database stands out for its precision and completeness, being one of the oldest wildfire databases in Europe, beginning in 1968 (Moreno et al., 2011; Vélez, 2001). Its inception coincided with the adoption in the same year of Law 81/1968 on Forest Fires, the first legal mandate expressly designed to address a serious problem. by means of prevention and control actions (López Santalla et al., 2017). The Bureau of Defense Against Forest Fires (Área de Defensa Contra Incendios Forestales: ADCIF) is the institution responsible for standardizing, maintaining, drafting and publishing these statistics, based on the information submitted by autonomous communities for every fire occurring in the country. All the baseline information collected is organized in different sections in the Spanish Forest Fire Reports (Parte de Incendio Forestal: PIF), which currently collects more than 150 data fields for each fire. It should be noted that this structure, sections and type of information gathered has varied over the years, undergoing a total of eight modifications from its first publication. Systematic collection of statistical data on forest fires began in 1956. Until then they were collected manually and on an irregular basis by the provincial services. In 1967, the Calculation Office of the Institute of Forestry Research and Experiences acquired a computer, which enabled a new model of PIF to be created that came into operation in the second semester of that year. Therefore, the first Annual Forest Fire Report was published in 1968, but included data on fires that had occurred since 1961. With regard to the quality of the data, it should be noted that, in the early years, it only included fires that affected forest masses or large non-forest areas, although subsequently, the rest of the fires were taken into account, even those of less than 1 ha. The spatialization of information has changed over the years from its beginning, when the minimum spatial unit of reference was the province (NUTS3), with a 10 x 10 km reference grid adopted after 1974. Until 1979, only those fires occurring in public and reforestation forests were recorded. Later, in the period 19801988, all fire events were collected, regardless of ownership. Since 1982, the municipality was added as a field in each fire location. Later, in 1989-1992, the PIF was reformed to incorporate important fields such as time, use of air, means or motivations related to intentionality. Since 1990, the General Statistic has been submitted to the European Commission for integration into the Community database EFFIS (European Forest Fire Information System). On the other hand, the traditional demarcation by region employed in the annual fire reports was the same for the period 1968-1977 with a total 7 regions (excluding the Balearic and Canary Islands): Galicia, North, Northeast, Ebro, Levante, the Hinterland and Andalusia. Since 1978, the number of regions increased to 10 (excluding the Canary Islands): Galicia, North or Cantabric (Asturias and Cantabria), Ebro (Aragón), Northeast (Catalonia-Baleares), Duero (Castilla-León), Center (Castilla–La Mancha), Levante (ValenciaMurcia), Extremadura, West Andalusia and East Andalusia. Moreover, from 1982 a more extensive section referring to weather conditions throughout the particular year was added, provided by the National Institute of Meteorology. From 1983, the previous regions were replaced by the Autonomous Communities. The sections of the current PIF contain the following common information: a) Location data: Includes the ID of the fire (IDPIF) as 10 digits. The codes of the autonomous community, province, municipality containing the fire ignition point (created in 1983), tile and grid (created in 1974) and UTM coordinates. Chapter 4: Materials and methods 24 b) Time data: Day, month, year, hour and minutes when the fire was detected, but also first arrival of engines by land (created in 1988), first arrival of fire-fighting aircraft (created in 1989), first airborne brigade arrival (created in 2005), and time when the fire was controlled and extinguished. c) Detection: Who first detected the fire (permanent guard, forestry officer, aircraft, etc.) and place of origin (road, path, house, train rail, crops, etc.). d) Ignition causes: Differences between known and supposed cause (since 1998), lightning, negligence and accidental causes, arson (created in 1989), unknown cause, rekindled fire (since 1998), identification or otherwise of the person causing it, and type of day (festival, Saturday, festival eve and working day). e) Danger conditions when the fire starts: Meteorological data (days from last rain, maximum temperature, relative humidity, wind) fuel model (since 1989) and probability of ignition. f) Type of fire: Surface, crown or subsoil (since 1989). g) Fire suppression media: Type of land transport (vehicles, helicopters), number of different personnel (technical staff, forestry agents, professional firefighters, civil staff, army, etc.) and extinguishing methods (aircrafts, helicopters, retardants, etc.). h) Fire suppression techniques: Direct or indirect attacks, firewall opening, etc. i) Losses: People killed and injured, civil protection incidents, type of surface affected, environmental impacts. 4.1.2. Fire data and fire features Fire features were retrieved from the General Wildfires Statistics (EGIF) database. Generally, fire records for 1974-2015 were selected and spatialized according to the 10 x 10 km UTM reference grid which is used by firefighting crews for approximate locations of fire ignition points. Fire count data, total burned area size, ignition triggering date and fire cause were retrieved for each event. In all cases, only information on fires larger than 1 ha was retained because small fires (i.e. fires with less than 1 ha affected) were not fully compiled until 1988. This is a well-known issue affecting other regions in the Mediterranean, such as Portugal (Pereira et al., 2011). Additionally, it is important to remember that in the autonomous community of Navarre, fire data were only available from 1988. Hence, all the analyses conducted in Navarre were based on a slightly different study period (from 1988 to 2010, 2013 or 2015). The start year was set as 1974, since it was the first year to use the 10 x 10 km grid. Prior to that time, fire data were only recorded at province level, so grid information was not available. The end year fluctuates depending on the temporal frame of other databases required for different analyses. For the first objective, the end year (2010) was chosen because of the availability of climate data from the MOTEDAS and MOPREDAS datasets (described below). For the second objective, the final year was set at 2013, because the sole input was the EGIF database, and at the time of the research, fire data was only available until then. As stated in section 1.2., regions were outlined following MAGRAMA specifications. In turn, two fire seasons were defined according to Moreno et al. (2014). Thus, annual data were divided into a springsummer season (S), from April to September; and an autumn-winter season (W) from October to March. From all available fire data information, several fire regime features were constructed separately for the season, region, NUTS3 and grid level. The final number of fire features changes according to each specific objective (see Table 2). Chapter 4: Materials and methods 25 Table 2. Summary of fire regime features constructed, their description and corresponding time period for each specific objective. Objective Fire Feature Description Time Period 1st. Explore the spatial-temporal distribution of fire regime features and their relation with climate-human factors Fire frequency (F) Total number of fires, regardless of size or ignition source 19742010 19742013 Burned area (B) Total fire affected area, regardless of size or ignition source Number of large fires (N500) Number of fires above 500 ha burned, regardless of ignition source Burned area from large fires (B500) Overall affected area from fires above 500 ha, regardless of ignition source Number of natural fires (NL) Number of fires triggered by lightning Burned area from natural fires (BL) Overall burned area from fires triggered by lightning Number of human fires (NH) Number of fires triggered by an anthropogenic source Burned area from human fires (BH) Overall burned area from fires triggered by an anthropogenic source 2nd. Estimate the contribution of fireweather danger on the temporal evolution of fire activity. Fire frequency (F) Total number of fires, regardless of size or ignition source 19792013 Burned area (B) Total fire affected area, regardless of size or ignition source 3rd. Analysis of spatial-temporal changes in the role of anthropogenic drivers on wildfires. Fire counts Number of fires by grid 19882010 19882013 Fire presence or absence Recoded into a binary presence or absence of fire recorded 25 subsets of occurrence Combination of two periods, two seasons, two causes and three fire sizes. 4th. Characterize the dynamics of recent-future fire regimes and know the drivers of their changes. Fire frequency (F) Total number of fires, regardless of size or ignition source 19742015 Burned area (BA) Total fire affected area, regardless of size or ignition source Burned area from natural fires (BAL) Overall burned area from fires triggered by lightning Chapter 4: Materials and methods 26 Burned area from large fires (BA100) Overall affected area from fires above 100 ha, regardless of ignition source Winter fire frequency (FW) Number of fires occurred in autumn-winter season, (W) regardless of size or ignition source In total, the calculation of fire occurrence (a total of 229,068 fires in the period 1974-2015, excluding small fires – i.e. less than 1 ha) was constructed by the method developed by De la Riva et al. (2004). This method consists in spatializing fire data as an input for fire modeling by using a kernel approach to interpolate historic fire observations. In terms of monthly mean and total values of the main fire features, a double annual peak can be found (the highest is usually found in August, with a second one in March, see Table 3). However, the second peak disappears for the area burned by large fires (>100 ha) and those caused by lightning, the latter showing a displacement of the summer peak to July. Table 3. Summary of monthly mean, standard deviation (sd) and total number of fires and burned area for each fire feature in the period 1974-2015 (small fires less than 1 ha are excluded). Fire feature Month Mean Sd Total Fire frequency January 0.05 188.33 7,199 February 0.13 479.82 18,691 March 0.24 754.77 33,329 April 0.12 428.94 17,311 May 0.06 135.09 7,921 June 0.07 163.06 9,892 July 0.17 387.86 24,226 August 0.34 675.74 47,657 September 0.29 808.65 40,296 October 0.09 398.99 13,160 November 0.03 103.22 3,979 December 0.04 167.94 5,407 Burned area January 2,477.85 3,278.01 104,069.51 February 6,133.04 8,977.37 257,587.55 March 11,321.46 11,995.31 475,501.25 April 6,277.16 7,694.76 263,640.90 May 2,985.79 4,329.85 125,403.29 June 5,778.87 9,068.18 242,712.68 July 28,265.16 37,616.06 1,187,136.62 August 44,804.75 34,226.26 1,881,799.50 September 25,413.33 35,333.18 1,067,360.03 October 7,011.48 12,163.58 294,482.01 November 1,865.65 3,731.95 78,357.49 December 3,024.84 5,501.75 127,043.30 Large burned area (> 100 has) January 1,004.03 1,775.17 42,169.42 February 2,349.64 4,567.64 98,684.74 March 3,859.71 4,371.90 162,107.73 Chapter 4: Materials and methods 27 April 2,623.81 4,282.10 110,199.93 May 1,520.87 3,105.52 63,876.42 June 4,033.21 8,588.28 169,394.96 July 23,179.12 35,556.11 973,523.08 August 34,094.29 27,675.81 1,431,960.15 September 16,189 25,104.63 679,938.17 October 4,087.58 7,829.04 171,678.17 November 1,071.2 3,344.38 44,990.17 December 1,795.32 3,886.1 75,403.49 Natural burned area January 0.39 1.51 16.50 February 5.01 22.38 210.30 March 38.91 226.98 1,634.38 April 20.30 53.86 852.43 May 49.33 93.17 2,071.79 June 593.73 1,194.33 24,936.83 July 5,772.42 16,780.78 242,441.65 August 2,469.47 4,578.96 103,717.77 September 523.36 1,081.92 21,980.98 October 16.06 46.77 674.57 November 5.89 30.26 247.40 December 0.22 0.76 9.30 On the other hand, several explanatory variables can be taken into account when addressing a fire regime characterization. These are usually divided into two groups: natural and human. In the first case, factors related to environmental conditions were selected to represent the general climate gradients and fireweather. The second group refers to anthropogenic conditions related with the fire ignition and were chosen on the basis of other previous research (Rodrigues et al 2014). 4.1.3. Climate and weather Climate data were extracted from MOTEDAS (Monthly Temperature Dataset of Spain) and MOPREDAS (Monthly Precipitation Dataset of Spain) datasets. These databases provide monthly climate information at a spatial resolution of 10 x 10 km. They were constructed from real measurements from the Spanish Meteorological Network of weather stations in the period 1951-2010 (González-Hidalgo et al., 2015, 2011). MOTEDAS and MOPREDAS stand out as one of the most accurate databases in the context of climate data for mainland Spain. Their development was based on the reconstruction of a meteorological data time series from each weather station in the region. This process includes a quality control, consisting of two steps: suspect data identification and inhomogeneity detection. Firstly, a set of reference series was calculated for each original station by means of a monthly correlation matrix between the candidate series and all the others, and selecting the neighboring series with the highest positive monthly correlation coefficient (mean greater than 0.60 and 0.50, for MOPREDAS and MOTEDAS, respectively) within a critical threshold distance of 25 km (for MOTEDAS) and 50 km (for MOPREDAS). The minimum overlapping period required for the correlation computation was set at 7 years from MOTEDAS and 10 years for MOPREDAS. Chapter 4: Materials and methods 28 To assess the suspect data, authors use both ratio and inter-quartile methods, as well as direct and inverse ratios to avoid the zero effect. On the other hand, for homogeneity analyses they applied a combination of tests: Single Normal Homogeneity Test -SNHT, Bivariate, t-Student and Pettit test. Finally, in order to fill the gaps in the data, the method consists in producing a combination of neighboring series with no overlapping periods. The final database of MOTEDAS consists of 3,066, and MOPREDAS of 2,670 selected homogenous series without suspect data from the different stations of AEMET (Agencia Estatal de Meteorología). After interpolating the stations’ data onto the 10 x 10 km grid cells, using an improved version comprising a combination of two weights: one radial weight with a Gaussian shape, and an angular weight. The radial weight prevents undesired exchange of information between different climatic regions, and between either side of the largest mountain chains. The angular weight avoids undesired overweighting of the areas with the highest station density. The mean number of stations involved in the estimation of each grid is approximately 4 for MOTEDAS and 6 for MOPREDAS. For the first objective of this dissertation, monthly data on annual average maximum temperature (T - Figure 4) and total precipitation in mm (P - Figure 5) in the period 1974-2010 were extracted and adapted to the fire grid using a nearest neighbor procedure. Both maximum temperature and precipitation were later reclassified into 10 homogeneous (equal interval) categories used to construct climate codes for the later fire features relationship plots. Figure 4. Spatial distribution of average maximum temperature (in ºC) from MOTEDAS. Chapter 4: Materials and methods 29 Figure 5. Spatial distribution of total precipitation (in mm) from MOPREDAS. The ERA-Interim Reanalysis datasets produced by the European Centre for Medium-Range Weather (ECMWF) dataset (Dee et al., 2011) was used to construct three different fire danger indexes (Fire Weather Index: FWI, US Burning Index: BI and Australian McArthur Forest Fire Danger Index: FFDI) necessary to achieve the 2rd objective. The main reason for this choice was due to the fact that this source has a higher spatial and temporal resolution (around 78 km). More specifically, 3-hourly 2 m air temperature, dew point temperature, surface total precipitation, and 10 m wind components were extracted to derive the following climate variables: maximum and minimum temperature, maximum and minimum relative humidity, maximum wind, total daily precipitation amount and total daily precipitation duration (see Jolly 2015 for more details). The WorldClim database is an interpolate climate surface for global land areas at a spatial resolution of 1 km (Hijmans et al., 2005). Monthly precipitation and mean, minimum, and maximum temperature were included as climate elements, and all input data came from different sources, restricted to all records for the 1950-2000 period. WorldClim was particularly chosen to create the Australian McArthur FFDI, providing the annual mean precipitation data, which when combined with the ECMWF maximum daily precipitation and temperature, produced the Drought Index (see Figure 2 in Chapter 6 and Jolly et al. (2015) for further details on the calculation process). It is therefore part of the achievement of the third objective of this dissertation. 4.1.4. Anthropogenic drivers The first of the human drivers refers to land use data, and was retrieved from Corine Land Cover 1990 (CLC), since it is centered on the study period. CLC information was used to outline the WildlandAgricultural Interface (WAI) and the Wildland-Urban Interface (WUI), two variables strongly related to anthropogenic ignitions (V. Leone et al., 2009; Martínez et al., 2004; Rodrigues et al., 2014). WAI represents the length of the boundary between agricultural and wildland areas, and WUI, the length between populated and wildland areas. Both were calculated at fire grid level (Rodrigues et al., 2016). Chapter 4: Materials and methods 30 On the other hand, in order to represent the human pressure over the wildlands, we have chosen the Demographic Potential (DP), which is an aggregate index for the ultimate future potential of the population, was retrieved from (J. L. Calvo and Pueyo, 2008) and based on the following formula: 𝑃𝑂𝑇𝑖= ∑( 𝑃𝑗 𝑑𝑖𝑗 2)+𝑃𝑖 𝑛 𝑗=1 ( 1 ) where POTi is the population potential accumulated in cell i, Pj are the inhabitants counted in each of the remaining accounting cells of the system and Pi are those of cell i itself, while d2 is the kilometre distance between each pair of cells i and j. In the cartographic values of POTi, those corresponding to its own resident population (Pi) plus those inferred by the rest of the system as a consequence of its positioning in the whole are accumulated, obtained by the sum of the population values of Pj divided by the distances (d) to which each accounting cell (j) is divided with respect to (i), and the latter elevated to an exponent, which in this case is 2, coinciding with the gravitational formula proposed by Newton. The demographic potential in 1991 was used at a spatial resolution of 5 x 5 km, later rescaled to the fire grid as the average value inside each cell (Figure 6). WAI, WUI (Figure 7 and Figure 8, respectively) and DP were normalized to a 0-1 interval and then aggregated to develop a Human Pressure Index (HPI, Figure 9), representing the overall pressure of human activities likely to result in fire ignition. Figure 6. Spatial distribution of the Demographic Potential (DP) in 1991. Chapter 4: Materials and methods 31 Figure 7. Spatial distribution of the wildland agricultural interface (WAI) length in meters. Figure 8. Spatial distribution of the wildland urban interface (WUI) length in meters. Chapter 4: Materials and methods 38 Remainder: the component that is left over from the two previous ones, and which therefore can be understood as anomalies or extreme events (both exceptionally high and low values) that are outside the average values of the trend and seasonal time series. Figure 12. Example of fire frequency time series decomposition in the Mediterranean region of mainland Spain. Autocorrelation Function (ACF) This is one of the simplest methods to check that a time series fulfills the characteristic of being stationary. Specifically, the idea is to observe whether every signal differs by a high degree of 0 for each time lag. With this purpose in mind, the ACF signal graph is visualized. In particular, a stationary signal produces few significant delays exceeding the ACF confidence interval. In comparison, another time series with a trend would show that, in most of its time lags, the confidence interval of the ACF is exceeded. Autoregressive Integrated and Moving Average (ARIMA) To forecast the evolution of fire features, a set of auto-regressive, integrated and moving average (ARIMA) models were employed. They can be viewed as a “filter” that tries to separate the signal from the noise, and the signal is then extrapolated into the future to obtain forecasts. Their main advantage is that they adjust exclusively to the historical series of the input variable, which greatly reduces the complexity of the analysis, since it is not necessary to incorporate other explanatory variables. However, the main condition of ARIMA models is that time series are stationary, i.e. constant in mean and variance. As this condition is very difficult to find, all fire feature time series were previously transformed through the square root and then a detransformation was applied to return to their original units. The future target period was set at 2016-2036, so that it would be the same length as the rest of periods. This exploration presupposed a continuous scenario in which it is assumed that the evolution of the factors associated with fire activity develop as observed in the whole historic period (1974-2015). Monthly time series of fire features for the current period (1995-2015) were entered into the ARIMA. The reason was to include the seasonal component (intra-annual peaks and drops) because they would offer more information to the model so that the future projection would be as consistent and realistic as possible. Chapter 4: Materials and methods 39 ARIMA offers several output data, the most important of which was the mean of the forecast, as well as the upper and lower limits of two confidence intervals (80% and 95%). An automatic ARIMA was applied to obtain future fire regime features, returning the best model according to the minimum Akaike information criterion (AIC) value, so its algorithm automatically calculates the p, i and q parameters. As reported by Hyndman and Khandakar (2008), the seasonal ARIMA formula is established as follows: Φ(𝐵𝑚)∅(𝐵)(1−𝐵𝑚)𝐷(1−𝐵)𝑑𝑦𝑡=𝑐+Θ(𝐵𝑚)𝜃(𝐵)𝜀𝑡 ( 6 ) where Φ(z) and Θ(z) are polynomials of orders P and Q respectively, each containing no roots inside the unit circle. If c≠0, there is an implied polynomial of order d + D in the forecast function. The main task in automatic ARIMA forecasting is selecting an appropriate model order, that is the values p, q, P, Q, D, d. When d and D are known, the rest of orders are chosen following an information criterion such as the AIC: AIC= −2log(𝐿)+2(𝑝+𝑞+𝑃+𝑄+𝑘) ( 7 ) where k=1 if 𝑐≠0 and 0 otherwise, and L is the maximized likelihood of the model fitted to the differenced data (1−𝐵𝑚)𝐷(1−𝐵)𝑑𝑦𝑡. The likelihood of the full model for 𝑦𝑡 is not actually defined and so the value of the AIC for different levels of differencing are not comparable. In order to overcome this difficulty, for our case of seasonal data, we selected the seasonally differenced data D =1. 4.2.3. Classification and regression Clustering The fire regime delimitation was done by Ward’s clustering method from the NbClust R package. This package provides a total of 30 indices for choosing the most adequate number of clusters and proposing the best clustering scheme from the different results obtained by varying all combinations of clusters (minimum and maximum desired), distance measurements and clustering methods. After several trial-anderror changing function parameters and according to the results obtained, the minimum and maximum number of clusters was established at 5 and 7, respectively. The distance selected was Canberra (Cd), and the method for the clustering outlined was Ward.D2. Cd was proposed by Lance and Williams (1967) and examines the sum of series of a fraction of differences between the coordinates of a pair of observations (Teknomo, 2015). In general, the Cd terms with zero numerator and denominator are omitted from the sum and treated as if the values were missing (Charrad et al., 2014). The formula of Cd is as follows: 𝐶𝑑 (𝑥,𝑦)=∑ |𝑥𝑗−𝑦𝑗| |𝑥𝑗|+|𝑦𝑗| 𝑑 𝑗=1 ( 8 ) where xj is the first observation with coordinates of the features and yj is the second observation with its corresponding coordinates of the same features. Each term of fraction difference has value between 0 and 1, although in itself it is not really between zero and one. If one of coordinates is zero, the term becomes 1 Chapter 4: Materials and methods 40 regardless of the other value, thus the distance will not be affected. Consequently, Cd is very sensitive to a small change when both coordinates area near to zero. Therefore, Cd has the advantage of not being affected by the presence of zeros, which are abundant in some cells of the study area, more especially in fire features such as natural burned area and large burned area (above 100 has). At each step the pair of clusters is chosen which leads to a minimum increase in the total within-cluster variance after merging. The Ward.D2 option implements Ward’s clustering criterion in which the dissimilarities are squared before clustering updating. K-Nearest Neighbor In order to transfer current fire regime clusters for the remaining time periods (past and future), a KNN classification was performed. KNN is a nonparametric technique used in statistical estimation and pattern recognition (Ripley, 1996) widely used since the 1970’s. The current period was taken as a benchmark, because it has more robust and reliable data. KNN trains for each grid in the test dataset (past and future fire features), finds the nearest K by a distance measure (Canberra distance), and the cluster class is decided by a majority vote of its neighbors. The K parameter means the maximum number of nearest neighbors considered in the algorithm (Venables and Ripley, 2002), being set in 5. Generalized Additive Models (GAM) In order to fulfil the first research objective for unravelling potential cause-and-effect relationships between fire features and climatic/human variables, several GAM regressions were calibrated for each Multidimensional Scatterplot (MDS) subset. Generalized Additive Models (GAM) are Generalized Linear Models (GLM) in which the usual linear relationships between the response and predictor variables are replaced by non-linear ‘smooths’ (Hastie and Tibshirani, 1986; Jones and Almond, 1992). The same as GLM, GAM can use probability distributions other than Gaussian, so we applied Negative Binomial to model the number of fires (N) and log linear distribution in burned area variables (B500, BL). NB is particularly suitable to deal with zero-inflated response variables, as is the case of N (Boadi et al., 2015). On the other hand, we applied a log linear family in burned area fire features (Hernandez et al., 2015). Model selection, is based on the reduction of Generalized cross validation (GCV, Craven and Wahba, 1978; Golub et al., 1979). GVC determines the optimal amount of smoothing and estimates the mean squared prediction error over all datasets where a single observation is omitted from the model fitting, and then predicted Deviance is explained (analogous to variance in a linear regression) and partial effects in the predictors were also calculated. All analyses for GAM modeling were conducted using the R package mgcv, version 1.8–9. Random Forest In order to assess the role of the drivers in fire regime change, we selected Random Forest (RF; Breiman, 2001) as the modeling algorithm, given its proven predictive accuracy (Bar Massada et al., 2011; Leuenberger et al., 2018a; Rodrigues and de la Riva, 2014a). RF is a tree-based ensemble algorithm that trains multiple decision trees by randomly bootstrapping the training sample, keeping 67% of the observations to train the decision tree and the remaining 33% (Out-of-bag, OOB) to evaluate the relative influence of the predictors and the model itself. The final stage assembles all trees into a final prediction as the average of all individual tree predictions (Bagging; Breiman, 2001). Chapter 4: Materials and methods 41 For each fire regime transition type, we trained and validated 100 RF models, using a random sample of 70% for training and the remaining 30% for testing the performance of the model. At the training stage, we conducted a 10-fold calibration procedure to identify the optimal parameters (mtry and ntrees) of the model. At the same time, we also evaluated the influence of each driver by calculating the percentage increase in the Mean Square Error (normalized between 1 and 0), and its explanatory sense by means of partial dependence plots (J.H. Friedman, 2001). To estimate the predictive performance of each model carried out, we calculated the Area Under the Receiver Operating Characteristic Curve (AUC; Bradley, 1997). Additionally, the explanatory sense of the covariates (either positively or negatively related) was explored by visual inspection of partial dependence plots. Geographically Weighted Regression Models (GWLR) GWR is a statistical technique for exploratory spatial data analysis developed within the framework of Local Spatial Models or Statistics. Local models could be described as the spatial disaggregation of global statistics whose main characteristic is that it is calibrated from a set of spatially limited samples and, hence, yielding local regression parameter estimates (Fotheringham et al., 2002). Therefore, GWR techniques extend the traditional use of global regression models, enabling local regression parameters to be calculated. Mathematically, a conventional GWR is described by the following equation: 𝑦𝑖= ∑𝛽𝑘 (𝑢𝑖 ,𝑣𝑖) 𝑋𝑘,𝑖 𝑘+ 𝜀𝑖 ( 9 ) where yi, xk,i, and 𝜀i are dependent variables, kth is the independent variable, and the Gaussian error at location i;(ui, vi) is the x–y coordinate of the ith location; and coefficients 𝛽 (ui, vi) are varying conditionals on the location. Such modeling is likely to attain higher performance than traditional regression models, and reading the coefficients can lead to a new interpretation of the phenomena under study. However, GWR models are not just a simple local regression model like, i.e., moving window regressions. In a moving window example, a region is drawn around a regression point and all the data points within this region (neighborhood) or window are then used to calibrate a model. This process is repeated over all the regression points, resulting in a set of local regression statistics. However, in this example, each point within the neighborhood is treated equally for regression purposes, no matter its distance to the target regression point. GWR overcomes this limitation by applying a distance weight pattern; hence, data points closer to the regression point are weighted more heavily in the local regression than data points farther away. In addition to the regression coefficients, a GWR model calculates several useful statistical parameters to analyze the spatial behavior of each explanatory variable, such as the value of the Student’s t test, which is used to determine the level of significance. On the other hand, GLM approaches such as Geographically Weighted Logistic Regression (GWLR) and Geographically Weighted Poisson Regression (GWPR) have been incorporated to GWR to extend its functionality (Fotheringham et al., 2002; Nakaya and Fotheringham, 2009). The GWR approach has been already been explored in several papers such as Koutsias, Martínez-Fernández, & Allgöwer (2010), Martínez-Fernández et al. (2013) and Rodrigues et al. (2014). These two methodologies—GWLR and GWPR—are used in this study to complement the results from GLM. Several parameters have been included when calibrating GWR models. Kernel shape and type, bandwidth selection and optimization parameters, or the local or global nature of the predictors (see Nakaya and Fotheringham, (2009) for further details of both method and software). In this project, GWR model fitting was carried out using Fixed Chapter 4: Materials and methods 42 Gaussian Kernel bandwidth, optimized according to the value of AICc, considering all the predictors as local covariates. 5 CHAPTER 5: SPATIALTEMPORAL DISTRIBUTION OF FIRE REGIME FEATURES This chapter describes the results, discussion and main conclusions obtained from the analyses related to the identification of the major fire regime features and their temporal dynamics. We evaluate the relationships of fire features with climate gradients and human pressure, as well as the assessment of the contribution of small fires into the fire regime characterization. Multi-Group Principal Component analysis, GAM models, change point detection methods, Mann-Kendall and Sen’ slope have been applied to fire features at regional and provincial level. The main goals are: describe and characterize the fire regime, identify its shifts and trends, determinate the extent to which fire regime is linked to climate-human factors and discover potential relations in the evolutions of fire features. Therefore, we seek to improve the understanding of the spatialseasonal patterns of the key fire regime features. Chapter 5: Spatial-temporal distribution of fire regime features 45 Chapter 5: Spatial-temporal distribution of fire regime features 46 Chapter 5: Spatial-temporal distribution of fire regime features 47 Chapter 5: Spatial-temporal distribution of fire regime features 54 Chapter 5: Spatial-temporal distribution of fire regime features 55 Chapter 5: Spatial-temporal distribution of fire regime features 56 Chapter 5: Spatial-temporal distribution of fire regime features 57 Chapter 5: Spatial-temporal distribution of fire regime features 58 Chapter 5: Spatial-temporal distribution of fire regime features 59 Chapter 5: Spatial-temporal distribution of fire regime features 60 Chapter 5: Spatial-temporal distribution of fire regime features 61 Chapter 5: Spatial-temporal distribution of fire regime features 62 Chapter 5: Spatial-temporal distribution of fire regime features 63 Chapter 5: Spatial-temporal distribution of fire regime features 70 Chapter 5: Spatial-temporal distribution of fire regime features 71 6 CHAPTER 6: THE INFLUENCE OF FIRE-WEATHER ON THE EVOLUTION OF FIRE ACTIVITY This chapter describes the results, discussion and main conclusions obtained from the analysis of spatial and temporal associations between monthly time series of fire weather danger indices (Fire Weather Index, Burning Index and Forest Fires Danger Index) at regional and local level. Decomposition of time series was the first step, then apply cross-correlation to explore seasonal associations at regional scale, as well as, a Pearson’s correlation was calculated between each index and 18 fire-activity subsets by fire size and cause at local scale. Chapter 6: The influence of fire-weather on the evolution of fire activity 75 Chapter 6: The influence of fire-weather on the evolution of fire activity 76 Chapter 6: The influence of fire-weather on the evolution of fire activity 77 Chapter 6: The influence of fire-weather on the evolution of fire activity 78 Chapter 6: The influence of fire-weather on the evolution of fire activity 79 Chapter 6: The influence of fire-weather on the evolution of fire activity 86 7 CHAPTER 7: CHANGE IN ANTHROPOGENIC DRIVERS This chapter describes the results, discussion and main conclusions obtained from the analyses of spatial and temporal evolution of human drivers factors into the fire regime features. We employed various regression models (Logit and Poisson Generalized Linear Models), as well as, trend analysis by means of Mann-Kendall. In addition, Geographically Weighted Regression Models are applied to assess spatial-temporal patterns. Chapter 7: Change in anthropogenic drivers 89 Chapter 7: Change in anthropogenic drivers 90 Chapter 7: Change in anthropogenic drivers 91 Chapter 7: Change in anthropogenic drivers 92 Chapter 7: Change in anthropogenic drivers 93 Chapter 7: Change in anthropogenic drivers 94 Chapter 7: Change in anthropogenic drivers 95 Chapter 7: Change in anthropogenic drivers 102 Chapter 7: Change in anthropogenic drivers 103 Chapter 7: Change in anthropogenic drivers 104 Chapter 7: Change in anthropogenic drivers 105 Chapter 7: Change in anthropogenic drivers 106 Chapter 7: Change in anthropogenic drivers 107 Chapter 7: Change in anthropogenic drivers 108 Chapter 7: Change in anthropogenic drivers 109 Chapter 7: Change in anthropogenic drivers 110 Chapter 7: Change in anthropogenic drivers 111 Chapter 7: Change in anthropogenic drivers 118 Chapter 7: Change in anthropogenic drivers 119 Chapter 7: Change in anthropogenic drivers 120 Chapter 7: Change in anthropogenic drivers 121 Chapter 7: Change in anthropogenic drivers 122 Chapter 7: Change in anthropogenic drivers 123 Chapter 7: Change in anthropogenic drivers 124 Chapter 7: Change in anthropogenic drivers 125 Chapter 7: Change in anthropogenic drivers 126 8 CHAPTER 8: EVOLUTION AND CAUSES OF FIRE REGIME CHANGE This chapter describes the main results, discussion and conclusions of the outlining of fire regime zones, their temporal evolution towards the near future and the analysis of the influence of drivers of fire activity in the observed fire regime trajectories. Random Forest is employed to evaluate the individual contribution of each fire driver, as well as, ARIMA models are used to forecast the immediate future trend of the main fire regime features. Appendix B 231 Appendix B 232 Appendix B 233 Appendix B 234 Appendix B 235 C APPENDIX C: SUPLEMANTARY MATERIAL OF DRIVERS OF CHANGE This appendix presents the supplementary material of the accepted paper entitled “Fire regime dynamics in mainland Spain. Part 1: drivers of change” which shows complementary results obtained in this publication. Appendix C 239 Appendix C 246 Appendix C 247 Appendix C 248 Appendix C 249 Appendix C 250 D APPENDIX D: PRELIMINARY PYROREGIONS DELIMITATION This appendix presents the work “Identifying pyroregions by means of Self Organizing Maps and hierarchical clustering algorithms in mainland Spain” forming part of the conference proceedings in Advances in Forest Fire Research 2018. It summarizes a preliminary attempt to define spatial-temporal definition pyroregions employing Self Organizing Maps (SOM), including the structural and trend component of fire regime features. 252 Appendix D 253 Appendix D 254 Appendix D 255 Appendix D 262 Appendix D 263 265 E APPENDIX E: CONFERENCES CONTRIBUTIONS This appendix brings together several abstracts from different conference contributions, mainly from the European Geosciences Union General Assembly (EGU) held in 2017 and 2018. Most of the abstracts refers to poster presentations and one to an oral dissertation. Appendix E 267 Appendix E 268 Appendix E 269 Appendix E 270