Next Article in Journal
Removing Moving Objects without Registration from 3D LiDAR Data Using Range Flow Coupled with IMU Measurements
Previous Article in Journal
Blind Adaptive Beamforming for a Global Navigation Satellite System Array Receiver Based on Direction Lock Loop
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 15
Issue 13
10.3390/rs15133388
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Deforestation as the Prominent Driver of the Intensifying Wildfire in Cambodia, Revealed through Geospatial Analysis

by
Min-Sung Sim
1,
Shi-Jun Wee
2,
Enner Alcantara
3 and
Edward Park
2,4,*
1
Department of Automotive Engineering, Yeungnam University, Gyeongsan-si 38541, Republic of Korea
2
Asian School of the Environment, Nanyang Technological University, Singapore 639798, Singapore
3
Institute of Science and Technology, São Paulo State University, São José dos Campos 01049, Brazil
4
Earth Observatory of Singapore, National Institute of Education, Nanyang Technological University, Singapore 639798, Singapore
*
Author to whom correspondence should be addressed.
Remote Sens. 2023, 15(13), 3388; https://doi.org/10.3390/rs15133388
Submission received: 18 May 2023 / Revised: 28 June 2023 / Accepted: 30 June 2023 / Published: 3 July 2023
(This article belongs to the Topic Application of Remote Sensing in Forest Fire)
Browse Figures
Review Reports Versions Notes

Abstract

:
Cambodia has the most fires per area in Southeast Asia, with fire activity have significantly increased since the early 2000s. Wildfire occurrences are multi-factorial in nature, and isolating the relative contribution of each driver remains a challenge. In this study, we quantify the relative importance of each driver of fire by analyzing annual spatial regression models of fire occurrence across Cambodia from 2003 to 2020. Our models demonstrated satisfactory performance, explaining 69 to 81% of the variance in fire occurrence. We found that deforestation was consistently the dominant driver of fire across 48 to 70% of the country throughout the study period. Although the influence of low precipitation on fires has increased in 2019 and 2020, the period is not long enough to establish any significant trends. During the study period, wind speed, elevation, and soil moisture had a slight influence of 6–20% without any clear trend, indicating that deforestation continues to be the main driver of fire. Our study improves the current understanding of the drivers of biomass fires across Cambodia, and the methodological framework developed here (quantitative decoupling of the drivers) has strong potential to be applied to other fire-prone areas around the world.

1. Introduction

The number of severe wildfire events has increased globally in recent decades, posing a multitude of environmental and health concerns [1]. Across the globe, the annual burnt area is close to 350 Mha, with CO2 emissions from wildfires amounting to over 50% of annual fossil fuel emissions, which threatens to set back efforts to combat climate change [2]. Thus, it is important to monitor the spatiotemporal patterns of wildfires as populations grow and communities expand [3]. The characteristics of fires vary across space, complicating the understanding of their drivers [4,5]. For instance, lightning is the main driver of fires in Canada’s boreal forest region [6]. In Mediterranean Europe, precipitation is strongly correlated with wildfires, as high precipitation increases fuel, such as vegetation, in the non-fire season and suppresses wildfires in the fire season [7]. Precipitation in Africa, with the savanna and desert climates, also has an important influence on wildfires, as dry weather triggers fires or drought reduces fuel loads [8,9]. In contrast, biomass fires in humid tropical and subtropical climates are often heavily influenced by human activities, such as agriculture and logging, and they seldom occur naturally [10,11].
While most biomass fires occur in Sub-Saharan Africa and South America globally, the tropical rainforests of Southeast Asia, home to over 680 million people, are emerging hotspots of wildfire activity. These events are highly critical in this region of over 680 million people due to the exposure of high population densities to its effects [12] and the fact that wildfires are historically a rare occurrence in these biomes [13]. According to a study by Vadrevu et al., (2019) [14], within Southeast Asia, Cambodia has the greatest number of detected fires per area, and fires in Cambodia showed a statically significant increase over the study period and the second-highest fire recurrence rate (repeat fires occurring in the same area) within Southeast Asia. In May 2016, more than 2000 ha of protected forests in the province of Battambang were burned, which affected biodiversity and crop yields, leading to extensive property damage [15]. Some studies have highlighted the close relationship between fires and agriculture in Cambodia, such as the use of fire for land clearance [16,17]. The country has lost more than 2.7 Mha of primary forest from 2001 to 2020 as a result of these activities, which accounts for more than a quarter of the country’s forested area [18,19]. Despite its frequent occurrence and the magnitude of environmental damages, wildfires in Cambodia are relatively understudied compared to Indonesia (its Southeast Asian counterpart) or other countries in the region. Even until now, the major drivers (i.e., environmental, climatic, or human) of wildfire occurrence in Cambodia remain poorly understood.
In this paper, we investigate the relative influence of each key driver on fire occurrences in Cambodia. We performed annual geospatial analyses (i.e., Geographically Weighted Regression) from 2003 to 2020, using publicly available remote sensing and meteorological data. The explanatory models incorporated a mix of climatic, topographic, and anthropogenic variables, and these quantify their relative influence on fire to better understand the spatiotemporal distribution of wildfires in Cambodia. Due to the nature of wildfires resulting from complex interactions between climate, terrain, and human activity, it is difficult to identify the drivers of wildfires without conducting in-depth studies at a national or regional scale [20]. The selection of drivers has a significant impact on the results of model predictions and is critical for interpreting the relationship between the drivers. There have been various approaches to analyzing the ignition and drivers of these fires, including simple indices, regression analysis, and machine learning-based modeling [21,22,23,24]. Statistical methods are used to establish a relationship between past fire data and fire drivers in order to identify the causes of forest fires and to perform spatial measurement, but modeling and prediction results are not accurate due to low learning ability and low error tolerance. Machine learning approaches can be more effective in interpretation by providing advanced learning capabilities, but they may overlook the spatial heterogeneity of geographical distributions due to their reliance on large amounts of data for learning, and statistical and machine learning methods assume spatial uniform relationships. To address this issue, a Geographic Weighted Regression (GWR) model has been introduced to enhance interpretation by analyzing the relationship between variables within a particular geographic area. A more precise comprehension of the spatial variability of wildfire drivers is crucial for effective wildfire analysis [25]. The novel approach developed in this study may be applied to better understand the relative influence of each driver of fire in other fire-prone regions around the world. Our study also has a strong potential to contribute to fire management for national/local authorities by identifying where each key driver of fire is dominant, as well as how these areas have been changing annually.

2. Data and Methods

2.1. Study Area and Spatial Regression Grids

Cambodia (181,035 square kilometers) is located within Mainland Southeast Asia (Figure 1A). The Mekong River flows to the east of the country, and in the center is Lake Tonlé Sap, Southeast Asia’s largest freshwater lake. Lake Tonlé Sap, the floodplain of the Mekong River, interacts with the Mekong River’s water level during the dry and rainy seasons. The country experiences a tropical rainforest climate with little variation in annual temperature, which ranges from 24 to 35 °C. The influence of monsoonal patterns has led to a distinct wet (May to October) and dry (November to April) season [26]. Calculating the annual fire count, the dry season accounts for more than 93 percent of the country’s wildfires, with a mean annual fire count of 31,000 during the study period (Figure 1B).
To facilitate spatial regression analysis, the study site was divided into square grids. We experimented with different grid sizes to obtain the best model fit: 5 × 5 km, 10 × 10 km, 15 × 15 km, 20 × 20 km, and 30 × 30 km. We selected 20 × 20 km grids, as they yielded a 5 to 10% improvement in the model fit compared to other grid scales, as of 2013. To further improve the model fit, we excluded incomplete grids (<400 km2) and grids with >30% surface area covered by water, including water bodies, such as Lake Tonlé Sap and the Mekong River. Water surface areas were calculated by overlapping the grids with a topographical raster (including water surface). A total of 315 grids remained for spatial regression analysis, although not all grids were utilized, as they may not contain any detected fire in a particular year.

2.2. Dependent Variable: Spatiotemporal Distribution of Fire

In this study, Moderate Resolution Imaging Spectroradiometer (MODIS) Active Fire data, obtained from NASA’s Fire Information for Resource Management System (FIRMS), were used to quantify the dependent variable [27]. MODIS, carried aboard the Terra and Aqua satellites, began collecting active fire data in 2000 and 2002, respectively. The data are collected by each satellite twice daily. The MODIS Active Fire Product records information, such as whether the fire was detected at day or at night, its location, and its detection confidence. Each observation is marked by a 1 km pixel, with its exact “location” at the center of the pixel. Since more than 93% of fires occur in the fire season, we only analyze fires that occur during the fire season from November to April.

2.3. Explanatory Variables: Identification, Filtering and Pre-Processing

We identified possible explanatory variables of fire, including meteorological, topographic, and human factors. The variables were selected based on whether their data were available throughout the study area and whether they were available on a monthly and/or annual basis. Nine explanatory variables were identified (Table 1): (1) Wind speed (WS), (2) Precipitation (PPT), (3) Maximum temperature (MT), (4) Soil moisture (SM), (5) Forest loss (FL), (6) Elevation (ELV), (7) Slope (SL), and (8) Population density (PD).
The meteorological variables were: wind speed, precipitation, maximum temperature, and soil moisture. Fuels with higher temperatures are more likely to reach ignition temperatures and combust more easily, contributing to their spread. Lower soil moisture increases flammability, thus increasing fire sensitivity in hot and dry climates [28,29,30]. Wind speed can influence fire spread rate [31]. Wind speed, precipitation, maximum temperature, and soil moisture data were obtained from TerraClimate [32], which provides monthly resolution data with global coverage, since Cambodian-specific data are not readily available. All meteorological variables were quantified using averaged data across the entire year, and their values were extracted and input into each grid.
Topographic variables: elevation and slope influence accessibility [33]. Areas with a lower elevation and flatter terrain are relatively more accessible and facilitate human activities, which increases the susceptibility of the land to fire [34]. Elevation also has an inverse relationship with temperature and oxygen level, ultimately influencing fire activity [35]. Elevation data were obtained from the Shuttle Radar Topography Mission Digital Elevation Model (SRTM DEM) data, which have spatial resolutions of 1 arc-second (30 m) [36]. Slope data were extracted from SRTM DEM using the Spatial Analysis function in ArcGIS. Afterwards, the average slope and elevation value per grid were calculated.
Human variables: we identified two proxies of anthropogenic activities: forest loss (or deforestation) and population density. Agricultural activity, involving slash-and-burn and deforestation, is known to contribute to biomass fires in Southeast Asia, including Cambodia [16,17,37]. Thus, we used population density data as a proxy of human activities contributing to fire activity. Data were obtained from WorldPop [38], which measures the number of people per square kilometer at a spatial resolution of 30 arc-seconds. The average population density (number of people per square km) was calculated per grid. Forest loss data were obtained from the Global Land Analysis & Discovery (GLAD) lab at the University of Maryland [39]. The product maps annual forest loss by identifying areas with stand-replacement disturbances or the removal of tree cover above 5 m, and it is available at a resolution of 30 m from 2003 to 2020. The proportion of forest loss area was tabulated per grid per study year.

2.4. Identifying Significant Drivers of Wildfire Using OLS

To decouple the drivers of Cambodia’s wildfires, we used a combination of Geographically Weighted Regression (GWR) and Ordinary Least Squares (OLS) regression [40,41]. Before performing GWR analysis, we first verified its applicability using OLS, which assumes that the relationship between a dependent and explanatory variable is constant across space. The OLS model is as follows:
Y = β 0 + ∑ j = 1 p β j χ j + ε
where Y is the dependent variable, β_0 is the model’s intercept, j is the jth explanatory variable (j = 1~p), and ε is the random error. As several explanatory variables had positively skewed distributions, we applied a log-transformation on these variables, so their values approximated a normal distribution. Afterwards, we performed OLS regression to analyze the linear relationship between each explanatory variable and fire count. Adjusted R2 and AICc were used to assess model performance, and the p-value was used to assess the statistical significance of each variable regarding fire count.
Out of the nine initially identified explanatory variables, we selected variables based on the following conditions: their (1) statistical significance (p ≤ 0.05) and their (2) degree(s) of multicollinearity with other explanatory variables, which was measured using the Variance Inflation Factor (VIF) [42,43]. We eliminated variables with VIF ≥ 2.5, considering them to have high multicollinearity. This process was iterated until we obtained a suitable OLS model using the appropriate explanatory variables. OLS analysis was conducted with data from 2013 and 2015, as they had the first- and second-most detected fires in Cambodia throughout the study period. Thus, the strength of the relationship between the explanatory and dependent variables are likely the best exhibited in these years.

2.5. Geographically Weighted Regression (GWR)

Unlike OLS, GWR accounts for spatial non-stationarity in the relationship between explanatory and dependent variables by allowing each variable’s regression coefficient to vary across space [44,45]. The model assigns weights to features used in each local regression, with the assigned weight decaying with distance from the regression point, and the GWR model’s equation is:
Y = β 0 u , υ + ∑ j = 1 p β j u , υ χ j + ε
where; Y represents the response (i.e., fire count), (u, v) represents the coordinates of the data point (i.e., grid), ε is the error according to an independent normal distribution with a mean of zero, and j is the number of explanatory variables used in the model. We performed GWR using ArcGIS, with its weighted function scheme constructed using Gaussian methods. To validate the GWR model, R2 and AICc values were used. First, residual maps of the OLS and GWR model were compared. In both models, residuals represent the discrepancy between observed and predicted values. This difference is an important indicator of the model’s ability to accurately predict the data. A positive residual indicates that the predicted value is higher than the observed value, and a negative residual value indicates that the predicted value is lower than the observed value. Residuals should ideally be randomly distributed and not exhibit spatial autocorrelation, which would indicate that the model has accounted for most key explanatory variables and is properly specified.
To assess the influence of each variable on fire count, a pseudo t-value was calculated per grid per variable by dividing the local coefficient with the standard error. Within each grid, the variable with the largest absolute pseudo t-value is deemed the most dominant and influential driver of fire. We analyzed the annual distribution of the most dominant drivers of fire by mapping pseudo t-values from 2003 to 2020. The methodological framework developed in this study is summarized in Figure 2.

3. Results

3.1. Analysis of OLS Regression and GWR for Model Validity

When all nine initially identified variables were input into an OLS model, the model explained 45% and 40% of the variance in fire count in 2013 and 2015, respectively. However, some explanatory variables had to be excluded due to significant multicollinearity with other variables, and/or they had low statistical significance (p > 0.05).
In the 2013 model, there were six statistically significant variables (p ≤ 0.05): wind speed, precipitation, maximum temperature, soil moisture, forest loss, and elevation. Among these variables, maximum temperature has a relatively high degree of multicollinearity (VIF ≥ 2.5). In the 2015 model, there were five statistically significant variables: wind speed, precipitation, soil moisture, forest loss, and elevation. As a result, only five variables were selected for further analysis: wind speed, precipitation, soil moisture, forest loss, and elevation.
When the OLS model was input with the five selected variables, the 2013 and 2015 models explained 52% and 47% of the variance in fire count with an improved model fit (corrected Akaike information criteria, or AICc = 662 and 710), respectively. All variables in 2013 and 2015 were statistically significant (p ≤ 0.05), and the VIF ranged from 1.11–1.44 (VIF < 2.5), indicating a low degree of multicollinearity among the variables (Table 2).
The residuals of the OLS model, using five variables (Figure 3A,B), were spatially autocorrelated, indicating that OLS is not a suitable approach, as it did not account for the nonstationary relationship between the explanatory and dependent variables.
The residuals of the GWR models were randomly distributed across the study area compared to those from the OLS model (Figure 3C,D). This indicates that the GWR model would be more appropriate than OLS, as the former has the ability to account for spatially non-stationary relationships, and the GWR model using the selected explanatory variables is properly specified. In addition, the GWR model had significant improvements over the OLS model. Comparing the performance of the GWR and OLS models using five explanatory variables, we found that the GWR model consistently had a higher adjusted R2 (i.e., 0.69–0.81 vs. 0.18–0.54) and a lower AICc (i.e., 302–474 vs. 540–811) than the OLS model for all years of analysis (Figure 3E). In addition, in the GWR model, R2 during the study period was relatively consistent.

3.2. Spatiotemporal Evolution of Fire Activity

Annual fire maps across Cambodia showed a clear clustering of the fire patterns (Figure 4). Fires were noticeably sparse in some areas, such as the southern plains and rice paddies near the capital, Phnom Penh, and the southeastern border with Vietnam. However, many fires were detected at forested areas in the east or the plains near Lake Tonlé Sap. Even denser fire activity was observed along the intersection between forests and cropland.
Throughout the study period, a subtle migration in grids with high fire activity from northwest of Lake Tonlé Sap to eastern Cambodia could be seen. The fires in western Cambodia, which were active from 2003 to 2011, have significantly decreased since 2012, and fires in eastern Cambodia have increased, especially from 2010 to 2016 (Figure 4).

3.3. The Relative Influence of Each Driver of Fire

By analyzing each variable’s local pseudo t-values, we found that deforestation had the strongest influence on fire (Figure 5A). The variable was most dominant across 59 ± 11% of the study area (and varied the least compared to other variables) throughout the study period. In particular, from 2010 to 2016, when the number of fires was high, and the number of fire-intensive areas increased, deforestation was more dominant in 67 ± 3% compared to other years. Precipitation was the second-most influential driver, dominant across 19 ± 13% of the study area, and dominated mainly in the southeast of Cambodia. The variable was especially influential in 2019 (32%) and 2020 (21%) in central and southeastern Cambodia compared to other years. Wind speed (18 ± 11%) was the third-most dominant driver for fire, mainly in the east and northeast. However, it was also influential in the southwest and the north. Elevation (13 ± 9%) and soil moisture (12 ± 6%) were, respectively, the second least and least influential drivers of fire, and their dominance was mainly limited to the areas around Lake Tonlé Sap (Figure 5B).
Overall, compared to deforestation, other variables indicated a very low count of dominance grids. Deforestation was consistently the most dominant driver of fire across the study area, evidenced by the lack of clear temporal trends in each variable’s dominance throughout the study period (Figure 5A).

4. Discussion

4.1. Model Validation for the Quantitative Analysis of Fire Drivers

Various steps taken, such as the removal of multicollinear variables through applied variance inflation factor thresholds (VIF ≤ 2.5) during the OLS stage, and analyzing the spatial distribution of the model residuals, help to minimize model design errors and improve the interpretability of the overall model results. We identified statistically significant explanatory variables and minimized multicollinearity by removing variables with VIF ≥ 2.5. Five variables were selected for analysis: wind speed, precipitation, soil moisture, forest loss, and elevation. Fire data were obtained from FIRMS’ fire product, which is widely used in fire analysis, along with Fire CCI. FIRMS’ fire data have the advantage of providing information on fire activity twice a day compared to Fire CCI, which generates monthly burned area products, enabling close-to-real fire tracking and monitoring. As OLS residuals and fire patterns exhibited considerable spatial autocorrelation, OLS was deemed an unsuitable approach for this study due to the nonstationary relationship between the explanatory and dependent variables. In contrast, GWR residuals were randomly distributed and did not exhibit severe spatial autocorrelation, indicating that this approach was superior to OLS when assessing spatially non-stationary relationships. Additionally, the GWR model had an improved performance over the OLS model, with consistently higher adjusted R2 (i.e., 0.69–0.81 vs. 0.18–0.54) and lower AICc (i.e., 302–474 vs. 540–811) values than its OLS counterpart (Figure 3E). Thus, GWR is highly effective in analyzing the spatially varying drivers of fire. In comparison, machine learning, also actively used to analyze and predict fire patterns [16], tends to overlook spatial heterogeneity and localized variations in the relationship between variables. In addition, it is usually more computationally intensive.

4.2. Dominant Drivers of Wildfire in Cambodia

We investigated five drivers of fire and found that deforestation was the most dominant driver. Overall, the influence of deforestation on fire did not change significantly throughout the study period (Figure 5A). The variable was especially dominant in areas with high fire activity (Figure 4), including provinces in the northwest (Battambang, Pailin) and east (Kartie, Kampong Thom) of Cambodia. The most active wildfire season is similar to the time when the dry season begins to prepare for the next agricultural activity. For example, this refers to burning agricultural residues or burning grass for reclamation [15,46]. This continues throughout the dry season and rarely happens as we enter the rainy season. Between 2012 and 2016, the number of fires increased significantly, possibly due to the increased deforestation rate since 2011 [47]. Before 2011, the annual deforestation rate was 0.9%, but since 2011, the annual deforestation rate has doubled to 1.8%. Most deforestation in the Pailin and Battambang regions of northwestern Cambodia occurred between 2003 and 2008 [48], consistent with the region’s very active fire period. There is a possibility that the fire may spread to other areas, depending on land use, including deforestation, such as the most densely concentrated fire zone, since clusters of fire activity post-2011 gradually shifted from the northwest to the east of the study area.
Precipitation was especially influential in southeast Cambodia (Figure 5B). We found that the influence of precipitation after 2019 (occupied 27 ± 5% of 20-km grids) was significantly higher than in previous years of the study (occupied 8 ± 2% of 20-km grids). We speculate that this increase may be attributed to recent, abnormally heavy rainfall and drought events that have rendered fires more sensitive to precipitation patterns. In particular, drought and flood damage occurred significantly in the southeastern region [49,50]. However, we are not certain if this was the actual reason, as the increase may simply be an anomaly, and two years of data are not sufficient to establish a significant trend. Longer-term studies of the relationship between rainfall and fire count are required to verify if the influence of rainfall on fire is increasing.
In the east of the study area, wind speed was usually the dominant driver of fire, although the size of the area where the variable was most influential was not consistent throughout the study period. The variable was also sporadically dominant in the center and south of the study area, and it is also challenging to identify any discernible trends in its influence. We are not able to pinpoint exactly why wind speed is the dominant driver of fire in these areas. It might be the result of other drivers having a relatively weaker influence. This can be seen as a loophole in which variables are not divided in more detail in complex interactions within environmental variables. In addition, soil moisture had a negative effect on fires around Lake Tonlé Sap, likely due to the influence of nearby floodplains and wetlands [29]. It is noteworthy that soil moisture was mainly dominant in areas with relatively less fires (Figure 4 and Figure 5B). High soil moisture lowers ambient temperature and inhibits fire propagation [51,52]. Elevation was dominant at the low elevation areas around Lake Tonlé Sap, likely due to ease of access to areas with lower elevation [35]. Similar to soil moisture, the variable is of low importance, as it is chiefly dominant in areas with less fires.
Our quantitative assessment of the drivers of wildfires in Cambodia revealed that areas with high fire activity were usually influenced by multiple drivers of fire. However, the complexity between fire and the various environmental factors limits our ability in quantifying the influence of drivers on fire. In particular, the two-way relationship between fire and deforestation makes it difficult to determine the extent to which fire is exclusive driven by deforestation activities, as tree cover loss may also be the result of biomass fires [39]. We emphasize the causal relationship between deforestation and fires in relation to the environmental characteristics and agricultural methods of the study area. In addition, numerous studies, i.e., Curtis et al., (2018) [53] and Tyukavina et al., (2022) [54] indicate that most tree cover loss in Cambodia and Southeast Asia is not caused by fire, but instead by non-fire deforestation. To discern the relationship between deforestation specifically and wildfires more accurately, it is crucial to conduct a quantitative analysis, distinguishing between overall forest loss and deforestation. This would require determining the proportion of forest loss that can be directly attributed to deforestation. In future studies, it will be necessary to clearly differentiate between deforestation and other forms of forest loss, as well as to explore, in greater detail, the underlying mechanisms by which deforestation influences wildfire occurrence and severity.

5. Summary

We assessed the relative influence of environmental, climatic, and anthropogenic factors on wildfires in Cambodia annually from 2003 to 2020. Using OLS, we identified statistically significant variables (p ≤ 0.05) and removed those that exhibited significant multicollinearity with other variables (VIF ≥ 2.5); the five variables chosen for further analysis were wind speed, precipitation, soil moisture, forest loss, and elevation. Next, we assessed the difference in OLS and GWR model performance using data from the two years in the study duration with the greatest fire activity in Cambodia (2013 and 2015). Using adjusted R2 and AICc values as metrics of comparison, we found that the GWR model had superior performance to the OLS model. The high autocorrelation observed in fire and OLS residuals also supported the use of GWR over OLS, as GWR was able to account for spatially non-stationary relationships. As a result of analyzing the pseudo-value of each variable, we found that deforestation was dominant in 48–70% of the study area and persisted as the most dominant driver during the study period. Other variables were less influential than deforestation and showed no clear trend. This supports our finding that deforestation is the main driver of fire throughout the study period. Deforestation was also usually dominant in areas with high fire activity, indicating its strong influence on fire activity relative to other drivers.
Our study contributes to the understanding of the drivers of wildfires in Cambodia, where national-level studies are relatively lacking compared to other countries in Southeast Asia, despite having the highest frequency magnitude of fire in the region. The novel approach developed in this study has strong potential to be applied to decouple and assess the drivers of fire in other regions around the world that likewise experience severe wildfires attributed to anthropogenic activities, such as land-use change.

Author Contributions

Conceptualization, E.P.; Methodology, M.-S.S.; Validation, E.P.; Investigation, M.-S.S.; Writing—original draft, M.-S.S.; Writing—review & editing, M.-S.S., S.-J.W., E.P. and E.A.; Project administration, E.P. All authors have read and agreed to the published version of the manuscript.

Funding

This study is funded by the Ministry of Education, Singapore, under its Academic Research Funds Tier 1 (#2021-T1-001-056 and #RG142/22 to EP), Tier2 (#MOE-T2EP402A20-0001 and #MOE-T2EP50222-0025) and the Earth Observatory of Singapore via its funding from the National Research Foundation Singapore under the Research Centres of Excellence Initiatives.

Data Availability Statement

The data that support the findings of this study are openly available in [DR-NTU] at (https://dr.ntu.edu.sg/ accessed on 17 May 2023).

Acknowledgments

The authors thank NASA Earth Data, MODIS fire datasets, and TerraClimate climate data. We wish to acknowledge the financial support for this study from the Ministry of Education of Singapore (#Tier2 MOE-T2EP402A20-0001). This study was supported by the Ministry of Education, Singapore, under its Academic Research Funds Tier 1 (#2021-T1-001-056 and #RG142/22 to EP), Tier2 576 (#MOE-T2EP402A20-0001 and #MOE-T2EP50222-0025) and the Earth Observatory of Singapore via its funding from the National Research Foundation Singapore under the Research Centres of Excellence Initiatives.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Coogan, S.C.; Robinne, F.; Jain, P.; Flannigan, M.D. Scientists’ warning on wildfire—A Canadian perspective. Can. J. For. Res. 2019, 49, 1015–1023. [Google Scholar] [CrossRef] [Green Version]
  2. Jolly, W.M.; Cochrane, M.A.; Freeborn, P.H.; Holden, Z.A.; Brown, T.J.; Williamson, G.J.; Bowman, D.M.J.S. Climate-induced variations in global wildfire danger from 1979 to 2013. Nat. Commun. 2015, 6, 7537. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  3. Wright, J.L.; Roy, S.S. Geospatial characteristics of Colorado wildfire occurrences from 2001 to 2020. Remote Sens. Appl. Soc. Environ. 2022, 28, 100872. [Google Scholar] [CrossRef]
  4. Govender, N.; Trollope, W.S.; Van Wilgen, B.W. The effect of fire season, fire frequency, rainfall and management on fire intensity in savanna vegetation in South Africa. J. Appl. Ecol. 2006, 43, 748–758. [Google Scholar] [CrossRef]
  5. Wan, C.; Roy, S.S. Geospatial characteristics of fire occurrences in southern hemispheric Africa and Madagascar during 2001–2020. J. For. Res. 2022, 34, 553–563. [Google Scholar] [CrossRef]
  6. Abdollahi, M.; Dewan, A.; Hassan, Q.K. Applicability of remote sensing-based vegetation water content in modeling lightning-caused Forest fire occurrences. ISPRS Int. J. Geo-Inf. 2019, 8, 143. [Google Scholar] [CrossRef] [Green Version]
  7. Oliveira, S.; Oehler, F.; San-Miguel-Ayanz, J.; Camia, A.; Pereira, J.M. Modeling spatial patterns of fire occurrence in Mediterranean Europe using Multiple Regression and Random Forest. For. Ecol. Manag. 2012, 275, 117–129. [Google Scholar] [CrossRef]
  8. De Sales, F.; Okin, G.S.; Xue, Y.; Dintwe, K. On the effects of wildfires on precipitation in Southern Africa. Clim. Dyn. 2019, 52, 951–967. [Google Scholar] [CrossRef] [Green Version]
  9. Wei, F.; Wang, S.; Fu, B.; Brandt, M.; Pan, N.; Wang, C.; Fensholt, R. Nonlinear dynamics of fires in Africa over recent decades controlled by precipitation. Glob. Chang. Biol. 2020, 26, 4495–4505. [Google Scholar] [CrossRef]
  10. Singh, M.; Sood, S.; Collins, C.M. Fire Dynamics of the Bolivian Amazon. Land 2022, 11, 1436. [Google Scholar] [CrossRef]
  11. Ying, L.; Han, J.; Du, Y.; Shen, Z. Forest fire characteristics in China: Spatial patterns and determinants with thresholds. For. Ecol. Manag. 2018, 424, 345–354. [Google Scholar] [CrossRef]
  12. Johnston, F.H.; Henderson, S.B.; Chen, Y.; Randerson, J.T.; Marlier, M.; DeFries, R.S.; Kinney, P.; Bowman, D.M.J.S.; Brauer, M. Estimated global mortality attributable to smoke from landscape fires. Environ. Health Perspect. 2012, 120, 695–701. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  13. Swaine, M.D. Characteristics of dry forest in West Africa and the influence of fire. J. Veg. Sci. 1992, 3, 365–374. [Google Scholar] [CrossRef]
  14. Vadrevu, K.P.; Lasko, K.; Giglio, L.; Schroeder, W.; Biswas, S.; Justice, C. Trends in vegetation fires in south and southeast Asian countries. Sci. Rep. 2019, 9, 7422. [Google Scholar] [CrossRef] [Green Version]
  15. Ministry Warns of Upcoming Dry Season That Can Spark Wildfires. Available online: https://www.khmertimeskh.com/501005100/ministry-warns-of-upcoming-dry-season-that-can-spark-wildfires/ (accessed on 17 May 2023).
  16. Jones, S.H. Vegetation fire and land use in Southeast Asia: The interpretation of remotely sensed data for Cambodia. Geocarto Int. 1998, 13, 63–73. [Google Scholar] [CrossRef]
  17. MacInnes, M. Governing the Flame: Bunong Management of Fire Regimes in Mondulkiri Province, Northeast Cambodia. In Proceedings of the FORTROP II Tropical Forestry Change in a Changing World, Bangkok, Thailand, 17–20 November 2008; pp. 124–131. [Google Scholar]
  18. Mermoz, S.; Bouvet, A.; Koleck, T.; Ballère, M.; Le Toan, T. Continuous Detection of Forest Loss in Vietnam, Laos, and Cambodia Using Sentinel-1 Data. Remote Sens. 2021, 13, 4877. [Google Scholar] [CrossRef]
  19. Rainey, H.; Bauran, H.; Evans, T. Forest Cover Trends in the Northern Plains of Cambodia 2002–2010; Wildlife Conservation Society: Phnom Penh, Cambodia, 2010. [Google Scholar]
  20. Thach, N.N.; Ngo, D.B.-T.; Xuan-Canh, P.; Hong-Thi, N.; Thi, B.H.; Nhat-Duc, H.; Dieu, T.B. Spatial pattern assessment of tropical forest fire danger at Thuan Chau area (Vietnam) using GIS-based advanced machine learning algorithms: A comparative study. Ecol. Inform. 2018, 46, 74–85. [Google Scholar] [CrossRef]
  21. Abid, F. A survey of machine learning algorithms based forest fires prediction and detection systems. Fire Technol. 2021, 57, 559–590. [Google Scholar] [CrossRef]
  22. Chang, Y.; Zhu, Z.; Bu, R.; Chen, H.; Feng, Y.; Li, Y.; Hu, Y.; Wang, Z. Predicting fire occurrence patterns with logistic regression in Heilongjiang Province, China. Landsc. Ecol. 2013, 28, 1989–2004. [Google Scholar] [CrossRef]
  23. Fujioka, F.M.; Gill, A.M.; Viegas, D.X.; Wotton, B.M. Fire danger and fire behavior modeling systems in Australia, Europe, and North America. Dev. Environ. Sci. 2008, 8, 471–497. [Google Scholar]
  24. Tonini, M.; Gill, A.M.; Viegas, D.X.; Wotton, B.M. A machine learning-based approach for wildfire susceptibility mapping. The case study of the Liguria region in Italy. Geosciences 2020, 10, 105. [Google Scholar] [CrossRef] [Green Version]
  25. Li, W.; Xu, Q.; Yi, J.; Liu, J. Predictive model of spatial scale of forest fire driving factors: A case study of Yunnan Province, China. Sci. Rep. 2022, 12, 19029. [Google Scholar] [CrossRef] [PubMed]
  26. Van Touch, R.M. Simulation Modelling of Alternative Strategies for Climate Change Adaptation in Rainfed Cropping Systems in North-Western Cambodia; Wrest Point Convention Centre: Hobart, Tasmania, 2015. [Google Scholar]
  27. Giglio, L.; Schroeder, W.; Justice, C.O. The collection 6 MODIS active fire detection algorithm and fire products. Remote Sens. Environ. 2016, 178, 31–41. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  28. Piñol, J.; Terradas, J.; Lloret, F. Climate warming, wildfire hazard, and wildfire occurrence in coastal eastern Spain. Clim. Chang. 1998, 38, 345–357. [Google Scholar] [CrossRef]
  29. Krueger, E.S.; Ochsner, T.E.; Engle, D.M.; Carlson, J.; Twidwell, D.; Fuhlendorf, S.D. Soil moisture affects growing-season wildfire size in the southern Great Plains. Soil Sci. Soc. Am. J. 2015, 79, 1567–1576. [Google Scholar] [CrossRef] [Green Version]
  30. Williams, A.P.; Abatzoglou, J.T.; Gershunov, A.; Guzman-Morales, J.; Bishop, D.A.; Balch, J.K.; Lettenmaier, D.P. Observed impacts of anthropogenic climate change on wildfire in California. Earth’s Future 2019, 7, 892–910. [Google Scholar] [CrossRef] [Green Version]
  31. Beer, T. The interaction of wind and fire. Bound.-Layer Meteorol. 1991, 54, 287–308. [Google Scholar] [CrossRef]
  32. Abatzoglou, J.T.; Dobrowski, S.Z.; Parks, S.A.; Hegewisch, K.C. TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958–2015. Sci. Data 2018, 5, 170191. [Google Scholar] [CrossRef] [Green Version]
  33. Sze, J.S.; Lee, J.S.H. Evaluating the social and environmental factors behind the 2015 extreme fire event in Sumatra, Indonesia. Environ. Res. Lett. 2019, 14, 015001. [Google Scholar] [CrossRef]
  34. Wharton, C.H. Man, fire and wild cattle in Southeast Asia. In Proceedings of the Annual Tall Timbers Fire Ecology Conference Series 8; Tall Timbers: Tallahassee, FL, USA, 1968. [Google Scholar]
  35. Satir, O.; Berberoglu, S.; Donmez, C. Mapping regional forest fire probability using artificial neural network model in a Mediterranean forest ecosystem. Geomat. Nat. Hazards Risk 2016, 7, 1645–1658. [Google Scholar] [CrossRef] [Green Version]
  36. Berry, P.; Garlick, J.D.; Smith, R.G. Near-global validation of the SRTM DEM using satellite radar altimetry. Remote Sens. Environ. 2007, 106, 17–27. [Google Scholar] [CrossRef]
  37. Yin, S. Biomass burning spatiotemporal variations over South and Southeast Asia. Environ. Int. 2020, 145, 106153. [Google Scholar] [CrossRef] [PubMed]
  38. Tatem, A.J. WorldPop, open data for spatial demography. Sci. Data 2017, 4, 170004. [Google Scholar] [CrossRef] [PubMed]
  39. Hansen, M.C.; Potapov, P.V.; Moore, R.; Hancher, M.; Turubanova, S.A.; Tyukavina, A.; Thau, D.; Stehman, S.V.; Goetz, S.J.; Loveland, T.R.; et al. High-resolution global maps of 21st-century forest cover change. Science 2013, 342, 850–853. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  40. Tulbure, M.G.; Wimberly, M.C.; Roy, D.P.; Henebry, G.M. Spatial and temporal heterogeneity of agricultural fires in the central United States in relation to land cover and land use. Landsc. Ecol. 2011, 26, 211–224. [Google Scholar] [CrossRef]
  41. Oliveira, S.; Pereira, J.M.C.; San-Miguel-Ayanz, J.; Lourenço, L. Exploring the spatial patterns of fire density in Southern Europe using Geographically Weighted Regression. Appl. Geogr. 2014, 51, 143–157. [Google Scholar] [CrossRef]
  42. Parajuli, A.; Gautam, A.P.; Sharma, S.P.; Bhujel, K.B.; Sharma, G.; Thapa, P.B.; Bist, B.S.; Poudel, S. Forest fire risk mapping using GIS and remote sensing in two major landscapes of Nepal. Geomat. Nat. Hazards Risk 2020, 11, 2569–2586. [Google Scholar] [CrossRef]
  43. O’brien, R.M. A caution regarding rules of thumb for variance inflation factors. Qual. Quant. 2007, 41, 673–690. [Google Scholar] [CrossRef]
  44. Fotheringham, A.S.; Brunsdon, C.; Charlton, M. Geographically Weighted Regression: The Analysis of Spatially Varying Relationships; John Wiley & Sons: Hoboken, NJ, USA, 2003. [Google Scholar]
  45. Rodrigues, M.; de la Riva, J.; Fotheringham, S. Modeling the spatial variation of the explanatory factors of human-caused wildfires in Spain using geographically weighted logistic regression. Appl. Geogr. 2014, 48, 52–63. [Google Scholar] [CrossRef]
  46. Fires-in-Cambodia. Available online: https://earthobservatory.nasa.gov/images/48859/fires-in-cambodia (accessed on 17 May 2023).
  47. Lohani, S.; Dilts, T.E.; Weisberg, P.J.; Null, S.E.; Hogan, Z.S. Rapidly accelerating deforestation in Cambodia’s Mekong River Basin: A comparative analysis of spatial patterns and drivers. Water 2020, 12, 2191. [Google Scholar] [CrossRef]
  48. Sourn, T.; Pok, S.; Chou, P.; Nut, N.; Theng, D.; Rath, P.; Reyes, M.R.; Prasad, P.V. Evaluation of land use and land cover change and its drivers in Battambang Province, Cambodia from 1998 to 2018. Sustainability 2021, 13, 11170. [Google Scholar] [CrossRef]
  49. Mekong Countries Urged to Take Bold Actions, as the River Experiences Three Consecutive Years of Low Flows. Available online: https://www.mrcmekong.org/news-and-events/news/pr-13-1-2022/ (accessed on 17 May 2023).
  50. Cambodia–Floods. Available online: https://reliefweb.int/report/cambodia/cambodia-floods-cambodia-meteo-aha-centre-media-echo-daily-flash-20-september-2019 (accessed on 17 May 2023).
  51. Valette, J.; Gomendy, V.; Maréchal, J.; Houssard, C.; Gillon, D. Heat-transfer in the soil during very low-intensity experimental fires-the role of duff and soil-moisture content. Int. J. Wildland Fire 1994, 4, 225–237. [Google Scholar] [CrossRef]
  52. Chaparro, D.; Vall-Llossera, M.; Piles, M.; Camps, A.; Rüdiger, C. Low soil moisture and high temperatures as indicators for forest fire occurrence and extent across the Iberian Peninsula. In Proceedings of the 2015 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Milan, Italy, 26–31 July 2015. [Google Scholar]
  53. Curtis, P.G.; Slay, C.M.; Harris, N.L.; Tyukavina, A.; Hansen, M.C. Classifying drivers of global forest loss. Science 2018, 361, 1108–1111. [Google Scholar] [CrossRef] [PubMed]
  54. Tyukavina, A.; Potapov, P.; Hansen, M.C.; Pickens, A.H.; Stehman, S.V.; Turubanova, S.; Parker, D.; Zalles, V.; Lima, A.; Kommareddy, I.; et al. Global trends of Forest loss due to fire from 2001 to 2019. Front. Remote Sens. 2022, 3, 825190. [Google Scholar] [CrossRef]
Remotesensing 15 03388 g001 550
Figure 1. (A) Overview of the study site and its major geographical features at the province level. Each red spot indicates the fire count at the location of each active fire pixel in 2013. (B) Annual fire count during the fire season (FS, in red) and non-fire season (NFS, in yellow) from 2003 to 2020. The horizontal dashed line indicates the mean annual fire count throughout the same time period. Mean monthly fire count from 2003 to 2020, with dry season months (November to April) represented in purple and wet season months (May to October) represented in gray.
Figure 1. (A) Overview of the study site and its major geographical features at the province level. Each red spot indicates the fire count at the location of each active fire pixel in 2013. (B) Annual fire count during the fire season (FS, in red) and non-fire season (NFS, in yellow) from 2003 to 2020. The horizontal dashed line indicates the mean annual fire count throughout the same time period. Mean monthly fire count from 2003 to 2020, with dry season months (November to April) represented in purple and wet season months (May to October) represented in gray.
Remotesensing 15 03388 g001
Remotesensing 15 03388 g002 550
Figure 2. The data processing and analysis workflow.
Figure 2. The data processing and analysis workflow.
Remotesensing 15 03388 g002
Remotesensing 15 03388 g003 550
Figure 3. Residual plot from the (A) 2013 and (B) 2015 OLS regression and the (C) 2013 and (D) 2015 GWR regression. (E) Comparison of GWR and OLS model performance per study year, with adjusted R2 and AICc as metrics of comparison.
Figure 3. Residual plot from the (A) 2013 and (B) 2015 OLS regression and the (C) 2013 and (D) 2015 GWR regression. (E) Comparison of GWR and OLS model performance per study year, with adjusted R2 and AICc as metrics of comparison.
Remotesensing 15 03388 g003
Remotesensing 15 03388 g004 550
Figure 4. The annual distribution and density of fire count per 20-km grid across Cambodia, from 2003–2020.
Figure 4. The annual distribution and density of fire count per 20-km grid across Cambodia, from 2003–2020.
Remotesensing 15 03388 g004
Remotesensing 15 03388 g005 550
Figure 5. (A) Number of 20-km grids where each driver was the most dominant driver of fire across the study area. (B) Annual distribution of the dominant drivers of fire.
Figure 5. (A) Number of 20-km grids where each driver was the most dominant driver of fire across the study area. (B) Annual distribution of the dominant drivers of fire.
Remotesensing 15 03388 g005
Table
Table 1. Description of the data in variables.
Table 1. Description of the data in variables.
CategoryVariableData SourceUnitResolution
Meteorological variablesWind speedTerraClimatem/s2003–2020
(Yearly)		4 km
Precipitationmm
Maximum Temperature°C
Soil moisturemm
Topographic variablesElevationSRTM DEMm2003–2020
(Yearly)		30 m
Slope°
Human variablesForest lossGlobal Land Analysis & Discovery (GLAD)%2003–2020
(Yearly)		30 m
Population densityWorldPop%2003–2020
(Yearly)		30 arc-seconds
(1 km)
Table
Table 2. Results of the OLS model using the five selected variables.
Table 2. Results of the OLS model using the five selected variables.
VariableStd Coef.StdErrort-Statisticp-ValueVIFStd Coef.StdErrort-Statisticp-ValueVIF
20132015
Wind speed0.1860.0092.0520.040 **1.22−0.1320.080−3.9080.000 **1.17
Precipitation−0.3670.111−5.6670.000 **1.18−0.1730.031−2.3850.017 **1.14
Soil moisture−0.1980.179−2.5090.012 **1.31−0.4210.179−2.3500.019 **1.44
Forest loss0.5240.05722.3200.000 **1.160.4050.1083.7260.000 **1.11
Elevation0.22400673.7740.000 **1.330.3350.0063.5260.001 **1.10
(**) indicates p ≤ 0.05 and VIF < 2.5.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Sim, M.-S.; Wee, S.-J.; Alcantara, E.; Park, E. Deforestation as the Prominent Driver of the Intensifying Wildfire in Cambodia, Revealed through Geospatial Analysis. Remote Sens. 2023, 15, 3388. https://doi.org/10.3390/rs15133388

AMA Style

Sim M-S, Wee S-J, Alcantara E, Park E. Deforestation as the Prominent Driver of the Intensifying Wildfire in Cambodia, Revealed through Geospatial Analysis. Remote Sensing. 2023; 15(13):3388. https://doi.org/10.3390/rs15133388

Chicago/Turabian Style

Sim, Min-Sung, Shi-Jun Wee, Enner Alcantara, and Edward Park. 2023. "Deforestation as the Prominent Driver of the Intensifying Wildfire in Cambodia, Revealed through Geospatial Analysis" Remote Sensing 15, no. 13: 3388. https://doi.org/10.3390/rs15133388

APA Style

Sim, M. -S., Wee, S. -J., Alcantara, E., & Park, E. (2023). Deforestation as the Prominent Driver of the Intensifying Wildfire in Cambodia, Revealed through Geospatial Analysis. Remote Sensing, 15(13), 3388. https://doi.org/10.3390/rs15133388

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop