Predicting forest fire kernel density at multiple scales with geographically weighted regression in Mexico

https://doi.org/10.1016/j.scitotenv.2020.137313Get rights and content

Highlights

  • Fire occurrence was predicted with GWR for kernel and regular grid density.

  • Human factors and aboveground vegetation carbon density influenced fire occurrence.

  • GWR model accuracy was higher for kernel density compared to regular grid.

  • Best compromise between model accuracy and spatial detail was obtained at 15–20 km.

Abstract

Identifying the relative importance of human and environmental drivers on fire occurrence in different regions and scales is critical for a sound fire management. Nevertheless, studies analyzing fire occurrence spatial patterns at multiple scales, covering the regional to national levels at multiple spatial resolutions, both in the fire occurrence drivers and in fire density, are very scarce. Furthermore, there is a scarcity of studies that analyze the spatial stationarity in the relationships of fire occurrence and its drivers at multiple scales.

The current study aimed at predicting the spatial patterns of fire occurrence at regional and national levels in Mexico, utilizing geographically weighted regression (GWR) to predict fire density, calculated with two different approaches –regular grid density and kernel density – at spatial resolutions from 5 to 50 km, both in the dependent and in the independent human and environmental candidate variables.

A better performance of GWR, both in goodness of fit and residual correlation reduction, was observed for prediction of kernel density as opposed to regular grid density. Our study is, to our best knowledge, the first study utilizing GWR to predict fire kernel density, and the first study to utilize GWR considering multiple scales, both in the dependent and independent variables. GWR models goodness of fit increased with fire kernel density search radius (bandwidths), but saturation in predictive capacity was apparent at 15–20 km for most regions. This suggests that this scale has a good potential for operational use in fire prevention and suppression decision-making as a compromise between predictive capability and spatial detail in fire occurrence predictions. This result might be a consequence of the specific spatial patterns of fire occurrence in Mexico and should be analyzed in future studies replicating this methodology elsewhere.

Introduction

The need to understand the spatial distribution of fire ignition and the relative importance of influencing environmental and human drivers is capital throughout the fire management decision-making processes (Schneider et al., 2008; Elia et al., 2019). It is also required to improve the effectiveness of strategic allocation of resources for fire prevention and suppression. It can help orient fuel treatment and pre-positioning of fire suppression crews, to reduce fire risk and associated environmental, economic and social impacts (e.g. Rodríguez y Silva et al., 2014; Curt et al., 2016).

Determining the key drivers of fire occurrence is critical for understanding spatial patterns of wildfire and implementing effective fire management (Camp and Krawchuk, 2017). Because of complex interactions and variations between areas of study, the relative importance of human and biophysical factors such as climate, topography and fuels in determining fire occurrence remains the subject of debate (e.g. Parisien et al., 2011; Fusco et al., 2016; Syphard et al., 2017). Some authors have proposed theories of fuel- and climatic- limitations on fire occurrence (Litell et al., 2009; Bradstock, 2010; Kahiu and Hanan, 2018), and these theories have been expanded to include human ignition limitations on a varying constraints hypothesis of fire occurrence (e.g. Krawchuk and Moritz, 2011; Camp and Krawchuk, 2017). Nevertheless, the relationships among fire occurrence and human and environmental drivers can largely vary between different ecosystems and regions (e.g. Parisien et al., 2016; Elia et al., 2019).

In addition to regional variations, the spatial characterization of wildfire distribution may be dependent on spatial scale, resulting from the interactions among top-down controls -such as climatic gradients and bottom-up controls -such as local fuel conditions, weather and topography- (Parisien et al., 2011; Falk et al., 2011; Liu et al., 2012). In spite of the importance of analyzing the effect of scale on the variations of relevant drivers of fire occurrence, the majority of studies on spatial fire occurrence prediction have focused on one scale only, ranging from local (e.g. Syphard et al., 2008; Guo et al., 2017) or regional analysis (e.g. Syphard et al., 2007; Su et al., 2019) up to national and global scales (e.g. Vasconcelos et al., 2001; Botequim et al., 2013; Chuvieco and Justice, 2010). In contrast, studies analyzing fire occurrence at multiple scales are still relatively scarce (e.g. Miranda et al., 2012; Wu et al., 2014; Romain et al., 2015). In particular, very few studies have been conducted at multiple scales integrating both the local/regional and the national/global scales (e.g. Parisien and Moritz, 2009; Koutsias et al., 2015).

The spatial patterns of fire density occurrence and its relationships with human and environmental drivers have been previously analyzed in the literature utilizing statistical techniques such as logistic regression (e.g. Vilar et al., 2010), classification and regression trees (Amatulli et al., 2006) or random forests (e.g. Wu et al., 2010; Oliveira et al., 2012). The spatial clustering of fires has been quantified in descriptive studies characterizing fire density spatial distributions (e.g. Podur et al., 2003; Hering et al., 2009; Fuentes-Santos et al., 2013). Some studies have furthermore quantified the influences of various controls that drive the intensity of fire density spatial patterns through spatial analysis techniques such as spatial point processes (e.g. Yang et al., 2008; Mundo et al., 2013). One of the limitations of many of the statistical techniques widely utilized for fire occurrence prediction is the assumption of coefficient stationarity –i.e. that the influence of different predictors of fire ignition is constant across space- (Fusco et al., 2016). However, research has shown that when large geographical study areas are involved, it would be more reasonable to find varied rather than constant relationships (e.g. Martínez-Fernández et al., 2013; Rodrígues et al., 2014; Nunes et al., 2016). Such stationary relationships can be better described with models that allow for local spatial variation of model coefficients, such as GWR, a relatively recent spatial analysis technique that is receiving increasing attention in the literature (e.g. Ávila-Flores et al., 2010; Sá et al., 2011; Oliveira et al., 2014; Guo et al., 2016; Rodrígues et al., 2018).

Studies utilizing GWR have focused on the prediction of fire occurrence, considering presence/absence of fire ignition points (e.g. Zhang et al., 2016; Guo et al., 2017; Rodrígues et al., 2014, Rodrígues et al., 2018) or on the prediction of fire density (e.g. Nunes et al., 2016; Koutsias et al., 2010; Su et al., 2019). Whereas point data are used to represent wildland fire ignition locations, surface data are generally utilized to represent human and environmental variables, with potential complications in relating one type of data to the other when performing a statistical comparison and analysis (Flowerdew and Pearce, 2001; De la Riva et al., 2004; Kuter et al., 2011), together with locational uncertainties in the ignition point position (e.g. Koutsias et al., 2004; Amatulli et al., 2007). The use of fire density, on the other hand, can overcome some of these limitations by representing the number of observed fires by a reference surface unit (e.g. Koutsias et al., 2004).

Previous studies utilizing GWR for fire density prediction have focused on the prediction of fire density at municipal or provincial levels (e.g. Nunes et al., 2016; Koutsias et al., 2010; Martínez-Fernández et al., 2013) or on the estimation fire density calculated as the number of ignitions or burnt area proportions at a pixel level surface, utilizing a fixed over imposed regular grid, with generally coarse pixel sizes (e.g. 55–10 km) for large study areas (e.g. Sá et al., 2011; Oliveira et al., 2014) and finer scales (1–5 km) for studies conducted at more regional studies (e.g. Su et al., 2019). Some studies have shown that the calculation fire density by superimposing a regular grid of quadrats over ignition points can be highly inconsistent depending on the magnitude of the positional errors and the resolution of the grid (Koutsias et al., 2004). In contrast, the technique of kernel density has been proposed to smooth the influence of the individual fire ignition point, addressing some of their inherent positional inaccuracies, and creating continuous density surfaces that can be used to delineate critical zones of fire occurrence at landscape level, at multiple scales as defined by the radius search or bandwidth (e.g. González-Olabarria et al., 2015; Koutsias et al., 2015).

While most of the previous literature on kernel density has focused on the bandwidth selection to identify scales that best describe the spatial patterns of fire density data (e.g. Kuter et al., 2011; González-Olabarria et al., 2015; Flores-Garnica and Macías-Muro, 2018), to our best knowledge, we are not aware of previous studies that have analyzed how the kernel bandwidth selection affects the correlation between fire occurrence and environmental or human variables at different scales. Furthermore, we are not aware of any study that has attempted to utilize GWR to predict the spatial patterns of fire occurrence, expressed as fire kernel density.

Some studies (e.g. Parisien and Moritz, 2009; Koutsias et al., 2015; Wei and Larsen, 2018) have highlighted the impact of the size of the area of study on the observed relationships between fire occurrence and its environmental and human drivers. The country of Mexico provides a good opportunity for fire occurrence spatial modeling, across varying regions, because of large variations in fuel, climatic, and environmental conditions (e.g. Vega-Nieva et al., 2018, Vega-Nieva et al., 2019; Briones-Herrera et al., 2019). While most of the previous studies in Mexico that have addressed the role of human and climatic variables on fire occurrence have focused on a local or regional scale (e.g. Muñoz et al., 2005; Drury and Veblen, 2008; Rodríguez-Trejo et al., 2008, Rodríguez-Trejo et al., 2011; Farfán et al., 2018; Cerano-Paredes et al., 2019), no previous study has attempted to predict the spatial patterns of fire occurrence from human and environmental variables at a national level in Mexico.

In summary, the need for and novelty of the current study can be synthesized as:

  • 1)

    There is a need for multiple scale studies that analyze the spatial variations in the role of fire occurrence drivers both between regions of varying human and environmental conditions and between different spatial resolutions from local/regional to national scales.

  • 2)

    To our best knowledge, no previous study has attempted to evaluate the performance of GWR to predict fire kernel density, in comparison with a regular grid density approach, as previously analyzed in the literature (e.g. Sá et al., 2011; Oliveira et al., 2014).

  • 3)

    Furthermore, the use of GWR to predict fire density at multiple spatial scales, both in the dependent and independent variables, has not –to our best knowledge-been previously tested.

The current study aimed at predicting and mapping the spatial distribution of fire density and the role of is causative drivers, at national and regional levels, at different spatial resolutions, in Mexico. In particular, the specific objectives of the current study were:

  • 1)

    To identify the main drivers of fire density, at different scales from 5 to 50 km, at a national level by cause and at a regional level in Mexico.

  • 2)

    To predict and map fire density and the role of its causative drivers at a national and regional levels in Mexico.

Section snippets

Study area

The study was conducted at two levels: 1) national, 2) regional. The area of study for the national analysis was the whole country of Mexico. Regions were defined based on Briones-Herrera et al. (2019), building on the North American level 3 ecoregions map (EPA, 2013), together with the consideration of spatial patterns of fire occurrence in the country from previous studies (Pompa-García et al., 2018; Vega-Nieva et al., 2018, Vega-Nieva et al., 2019). Regions were updated to the most recent

National level

For the human caused fires, road and agricultural density, together with aboveground carbon density, were consistently selected for prediction of fire density at all investigated scales (Table 2). The resolution of these variables, however, varied with the scale of the fire density. Finer scale fire densities (5 and 10 km) were best predicted with finer resolutions (25 km and 5 km) for road density and agricultural interface density, respectively. Coarser resolutions in the predictor variables,

National level

The observed variations in the role of road density, with mainly positive relationships with fire occurrence for human fires and mainly negative relationships for the natural fires, confirm a strong effect of human accessibility in human caused fires in Mexico. These results agree with studies that have found a more important role of human variables of landscape accessibility in human compared to natural caused fires elsewhere (e.g. Narayanaraj and Wimberly, 2012; Liu et al., 2012; Wu et al.,

Conclusions

The current study is, to our best knowledge, the first analysis utilizing geographically weighted regression to understand the spatially varying role of human and environmental drivers of fire occurrence considering multiple scales, both in the dependent and independent variables.

Both at national and regional levels, human variables, followed by aboveground carbon density, were selected as the main drivers of fire occurrence for all the regions, fire causes and scales analyzed. Our observed

CRediT authorship contribution statement

Norma Angélica Monjarás-Vega: Conceptualization, Formal analysis, Writing - original draft. Carlos Ivan Briones-Herrera: Formal analysis. Daniel José Vega-Nieva: Conceptualization, Methodology, Formal analysis, Writing - original draft. Eric Calleros-Flores: Data curation. José Javier Corral-Rivas: Methodology, Supervision. Pablito Marcelo López-Serrano: Methodology, Supervision. Marín Pompa-García: Writing - review & editing. Dante Arturo Rodríguez-Trejo: Writing - review & editing. Artemio

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

Funding for this work was provided by CONAFOR/CONACYT Projects “CO2-2014-3-252620” and “CO-2018-2-A3-S-131553” for the development and enhancement of a Forest Fire Danger Prediction System for Mexico, funded by the Sectorial Fund for forest research, development and technological innovation “Fondo Sectorial para la investigación, el desarrollo y la innovación tecnológica forestal”.

We would like to thank CONAFOR's personnel for their support to the current study and for providing the fire

References (100)

  • S. Oliveira et al.

    Exploring the spatial patterns of fire density in Southern Europe using geographically weighted regression

    Appl. Geogr.

    (2014)
  • J. Podur et al.

    Spatial patterns of lightning-caused forest fires in Ontario, 1976-1998

    Ecol. Model.

    (2003)
  • M. Rodrígues et al.

    Modeling the spatial variation of the explanatory factors of human-caused wildfires in Spain using geographically weighted logistic reression

    Appl. Geogr.

    (2014)
  • P. Schneider et al.

    A VARI-based relative greenness from MODIS data for computing the fire potential index

    Remote Sens. Environ.

    (2008)
  • Z. Wu et al.

    Relative effects of climatic and local factors on fire occurrence in boreal forest landscape of northeastern China

    Sci. Total Environ.

    (2014)
  • A.A. Ager et al.

    Wildfire risk estimation in the Mediterranean area

    Environmetrics

    (2014)
  • M.A. Amaral-Turkman et al.

    Hierarchical space-time models for fire ignition and percentage of land burned by wildfires

    Environ. Ecol. Stat.

    (2011)
  • G. Amatulli et al.

    Assessing long-term fire risk at local scale by means of decision tree technique

    J. Geophys. Res.

    (2006)
  • S. Archibald et al.

    What limits fire? An examination of drivers of burnt area in Southern Africa

    Glob. Chang. Biol.

    (2009)
  • D. Ávila-Flores et al.

    Driving factors for forest fire occurrence in Durango State of Mexico: a geospatial perspective

    Chinese Goegraphical Science

    (2010)
  • B. Botequim et al.

    Developing wildfire risk probability models for Eucalyptus globulus stands in Portugal

    iForest

    (2013)
  • R.A. Bradstock

    A biogeographic model of fire regimes in Australia: current and future implications

    Glob. Ecol. Biogeogr.

    (2010)
  • C.I. Briones-Herrera et al.

    Modeling and mapping forest fire occurrence from aboveground carbon density in Mexico

    Forests

    (2019)
  • G. Caccamo et al.

    Using remotely-sensed fuel connectivity patterns as a tool for fire danger monitoring

    Geophys. Res. Lett.

    (2012)
  • P.E. Camp et al.

    Spatially varying constraints of human-caused fire occurrence in British Columbia, Canada

    Int. J. Wildland Fire

    (2017)
  • G.R.L. Carrillo et al.

    Análisis espacial de peligro de incendios forestales en Puebla, Mexico

    Interciencia

    (2012)
  • O. Cartus et al.

    A national, detailed map of forest aboveground carbon stocks in Mexico

    Remote Sens.

    (2014)
  • J. Cerano-Paredes et al.

    Climatic influence on fire regime (1700 to 2008) in the Nazas watershed, Durango, Mexico

    Fire Ecology

    (2019)
  • E. Chuvieco et al.

    Relations between human factors and global fire activity

  • A.D. Cliff et al.

    Spatial Processes: Models and Applications

    (1981)
  • CONAFOR (Comisión Nacional Forestal)

    Serie histórica anual de incendios

  • S. Costafreda-Aumedes et al.

    Human-caused fire occurrence modelling in perspective: a review

    Int. J. Wildland Fire

    (2017)
  • CRAN

    Moran.I. Moran's I autocorrelation index. R package

  • T. Curt et al.

    Modeling the spatial patterns of ignition causes and fire regime features in south France: implications for fire prevention policy

    Int. J. Wildland Fire

    (2016)
  • S.A. Drury et al.

    Spatial and temporal variability in fire occurrence within the Las Bayas Forestry Reserve, Durango, Mexico

    Plant Ecol.

    (2008)
  • M. Elia et al.

    Modeling fire ignition patterns in Mediterranean urban interfaces

    Stoch. Env. Res. Risk A.

    (2019)
  • EPA (United States Environmental Protection Agency)

    Level III Ecoregions of the Conterminous United States

    (2013)
  • ESRI

    ArcGIS V. 10.5 Software by Esri®

    (2016)
  • D.A. Falk et al.

    Multi-scale controls of historical forest-fire regime: new insights from fire-scar networks

    Front. Ecol. Environ.

    (2011)
  • G.M. Farfán et al.

    Modeling antropic factors as driver of wildfire occurrence at the Monarch Butterflu Biosphere

    Maderas y Bosques

    (2018)
  • J.G. Flores-Garnica et al.

    Bandwith selection for kernel density estimation of forest fires

    Revista Chapingo Serie Ciencias Forestales y del Ambiente

    (2018)
  • R. Flowerdew et al.

    Linking point and area data to model primary school performance indicators

    Geogr. Environ. Model.

    (2001)
  • E.J. Fusco et al.

    Quantifying the human influence on fire ignition across the western USA

    Ecol. Appl.

    (2016)
  • J.L. Gittleman et al.

    Adaptation: statistics and a null model for estimating phylogenetic effects

    Syst. Zool.

    (1990)
  • J.R. González-Olabarria et al.

    Different factors for different causes: analysis of the spatial aggregations of fire ignitions in Catalonia (Spain)

    Risk Anal.

    (2015)
  • F. Guo et al.

    Geospatial information on geographical and human factors improved anthropogenic fire occurrence modeling in the Chinese boreal forest

    Can. J. For. Res.

    (2016)
  • F. Guo et al.

    Spatial modelling of fire drivers in urban-forest ecosystems in China

    Forests

    (2017)
  • T. Hastie et al.

    Generalized Additive Models

  • A.S. Hering et al.

    Modeling spatiotemporal wildfire ignition point patterns

    Environ. Ecol. Stat.

    (2009)
  • INEGI (Instituto Nacional de Estadística y Geografía-México)

    Guía Para la Interpretación de Cartografía: Uso del suelo y Vegetación. escala 1, 250, 000: Serie V

    (2011)
  • Cited by (43)

    • A novel fire danger rating model based on time fading precipitation model — A case study of Northeast China

      2022, Ecological Informatics
      Citation Excerpt :

      Some models convert discrete fire points into continuous fire point density. Both Monjarás-Vega et al. (2020) and Oliveira et al. (2012) use kernel density functions to calculate forest fire density data. Monjarás-Vega et al. utilizing geographically weighted regression (GWR) to predict fire density.

    View all citing articles on Scopus
    1

    Contributed equally to this study.

    View full text