Modeling Seasonal Fire Probability in Thailand: A Machine Learning Approach Using Multiyear Remote Sensing Data

1. Introduction

Thailand experiences seasonal forest and agricultural fires beginning in January or February and continuing until the onset of the rainy season in May [1,2,3,4,5]. Predicting seasonal wildfires in this time period is critical to mitigating the negative environmental and human health impacts, including landscape degradation and high levels of hazardous PM 2.5 air pollution [1,6,7,8,9,10,11].

Wildfires in Thailand are challenging to predict as they are largely driven by anthropogenic, rather than natural factors, including cultural, economic, and legal activities [12,13]. These anthropogenic influences not only drive current fire events but also initiate feedback mechanisms that create conditions conducive to future wildfires (Figure 1) [14,15,16,17,18]. More specifically, wildfires most commonly result from escaped intentional burns, carelessness, or arson, though some are ignited by lightning [1,17]. Intentional burns are used for clearing vegetation for agriculture and hunting, clearing understory for non-timber forest product (NTFP) collection, eliminating crop residue from rice, maize, and sugarcane plantations, and regenerating grass on livestock grazing land [15,16,18,19,20,21,22,23,24]. Only 5% of forest fires between 2001-2021 in Thailand were associated with complete forest loss, suggesting that forests are likely degraded but not completely cleared by most types of fire events [2].

Fire feedback mechanisms are both natural and anthropogenic. Certain forest types, such as dry dipterocarp, bamboo, and teak forests, are naturally more fire-prone and are also more frequently burned for NTFP collection [16,17]. Many of these ecosystems are maintained naturally by edaphic, climatic, and ecological factors [25,26,27,28] while others are likely the result of human-facilitated encroachment on other ecosystems [25,26,29,30] and may eventually return to wet forest types if left unburned [26,31]. The severity of wildfires is seasonally influenced by El Niño and La Niña [32] and is continually exacerbated by population growth and the growing need for land for cultivation [23,33,34].

Together, these complexities make it difficult to accurately predict and manage fires in Thailand. Mapping wildfire probability and risk are critical components of wildfire management. Fire risk is the combination of fire probability, intensity, exposure, and susceptibility [35]. Fire risk maps predict where fires will occur and how dangerous they will be [35,36,37]. In this study, we will refer to “fire probability” to mean the likelihood of a fire occurring and “fire risk” to mean the comprehensive quantification of probability, intensity, exposure, and vulnerability. Note that in the Thai language scientific and government literature frequently uses “fire risk” to refer to what English speaking fire managers define as “fire probability”. This study addresses fire probability, which is foundational to quantifying fire risk but does not provide a complete assessment of fire risk.

Existing work on fire probability mapping in Thailand and peninsular Southeast Asia includes operational methods and scientific studies (Supplement Tables S1 and S2). Operational methods in Thailand generally use multi-criteria analysis (MCA) to evaluate fire probability, which is a method that weights predictor variables based on expert opinion or statistical distributions and combines them into a final score [1,10,38]. Scientific literature exploring fire probability mapping in peninsular Southeast Asia uses either MCA [39,40,41,42,43], probabilistic statistics (PS) [44,45], or machine learning (ML) [46,47,48,49,50,51], approaches, with MCA and ML being the most common. The most common MCA approach is Analytical Hierarchy Process (AHP), while many different ML approaches are used, including Random Forest (RF), Bayes Network (BN), Naive Bayes (NB), and Support Vector Machine (SVM).

While providing important information for wildfire management, existing approaches for fire probability mapping face a few key challenges related to methods and data. With regards to methods, the MCA approaches are the simplest to implement, however scores and weights are completely subjective and final scores can be difficult to compare to one another or interpret in real world scenarios [50,52,53,54,55,56,57]. Existing MCA approaches have seen limited operational use by the general public in Thailand, due in part to how difficult they are to interpret in a practical setting without a fundamental understanding of the analysis methods [14]. ML approaches are computationally expensive and require large reference data sets for training and validation [58,59]. However, they can be used to empirically examine the relationship between predictor variables and fire occurrence and output probabilities that can be easily interpreted as the likelihood of a real world event [55,58,59]. In addition, ML based studies provide standard measures of accuracy and empirical measures of variable importances, allowing for inter-model comparison [53,54,55,57,59].

Existing approaches also do not include data-driven temporal variation between the predictor variables and predicted fire probabilities. Predictor variables commonly include burn history and frequency, land use and land cover, vegetation and water indices, fire season weather, infrastructure and accessibility, and topography [1,10,38,39,40,41,42,43,44,45,46,47,48,49,50,51]. Many of the approaches produce a single static map of fire probability representative of the entire time period of interest; these probabilities are based on cumulative fire occurrence over a set of years, and average conditions from the same years or a subset of those years [39,40,42,43,44,45,46,47,48,49,50,51]. They use historical fire reference points from a range of years, but most predictor variables are static; any historical predictor variables available for the time period are aggregated into a single layer for use in the model, rather than associating the predictor variables to reference points of each corresponding year [39,40,42,44,45,46,47,48,49,50,51]. Further, many studies derive their fire occurrence data solely from MODIS or VIIRS active fire hotspots [44,45,46,50,60]. The inadequate spatial and temporal resolution and sensor type make these datasets ineffective in detecting the small, short, and mostly anthropogenic fires common in Thailand.

To address both methodological and temporal data issues, we developed a seasonal fire probability mapping approach for Thailand, with a fully developed example from the nine northernmost provinces of the country. We leverage machine learning and globally available data sets in Google Earth Engine’s (GEE) cloud computing environment. We trained a random forest model on annual historical fire and fire predictor data from 2016 to 2023 in Northern Thailand, then deployed the model on 2024 fire predictor data to generate the 2024 fire probability map. This methodology can increase the speed and complexity of fire risk related analyses while also offering flexibility for the input data sources and analysis frequency; it can be easily run for any given year for which historical fire data and predictor data are available.

Our intention was to develop a method to produce a straightforward probability metric that is easily interpreted by the general public and can be operationalized by government fire managers by running the analysis annually before each fire season. The approach is easy for local fire experts to modify using higher quality local data in lieu of publicly available global datasets used by default. Our approach is novel in that we offer an operationalization-ready machine learning approach for fire probability mapping, which improves upon the current operational methods in Thailand that use manual weighted overlay analyses; we use historical data to account for interannual variation in fire and its predictors in order to predict seasonal fire probability in a given year; we use a more representative sample design that allows for locations with any burn history to be selected as fire presences and absences for a given year; we empirically test and refine a comprehensive set of potential fire predictors compiled from existing fire probability mapping methodologies; and we provide all code necessary for implementation.

2. Materials and Methods

2.1. Study Area

Our study area was the 9 most northeastern provinces of Thailand: Chiang Mai, Chiang Rai, Mae Hong Son, Lamphun, Lampang, Phayao, Phrae, Nan, Tak (Figure 2). These provinces experience lengthy fire seasons [2,41,50] characterized by infrequent, small, fragmented, and low intensity fires in forested and agricultural lands [2]. In 2019, the LULC was comprised of 69% closed forest (>70% canopy cover); 16% agriculture; 11% open forest (15-70% canopy cover); 2% urban; and <1% grassland, shrubland, or herbaceous wetland (<10% canopy cover) [61]. The terrain is mountainous, with some of the highest elevations found in Thailand [62]. The three main ecoregions are Kayah-Karen montane rain forests, Central Indochina dry forests, and Northern Thailand-Laos moist deciduous forests [63], all tropical and subtropical moist broadleaf forest biomes. The region is characterized by strong seasonality in precipitation and vegetation greenness; precipitation peaks in September through October and the normalized vegetation index (NDVI) peaks in July through October [64,65,66,67]. Mean annual precipitation accumulation for 2016-2024 was 163 cm and mean temperature for 2016-2024 was 24 degrees celsius [68]. Temperature generally fluctuates between 20-30 degrees celsius throughout the year with the lowest temperatures in November through January and the highest temperatures in March through May [64].

2.2. Model Selection

We chose a random forest model to model fire probability due to performance and operationalizability (Figure 3) [69]. Random forest models have performed well in predicting fire likelihood in Thailand and neighboring countries, sometimes outperforming more complex machine learning models [46,50]. They better capture non-linear interactions between predictors [70] and are more robust to overfitting due to data noise [71,72], and multicollinearity [73,74], than simpler decision tree and regression methods. In addition, random forest models are easy to operationalize as they are conceptually straightforward algorithms that can be easily understood, employed, and refined by fire managers, even with no prior machine learning experience.

To create the seasonal fire probability map, we used Google Earth Engine’s smileRandomForest classification algorithm with 500 decision trees [58,75]. Training points consisted of fire presence and absence data, and environmental and social predictor variables consisted of representative data layers. Once trained, the model was deployed on predictor variable layers from 2024 to produce the probability map for the 2024 fire season. The model outputs probability values that can be interpreted as the percentage of decision trees that predict fire presence [58].

2.3. Data Selection

2.3.1. Fire Presence and Absence Data

Reference points representing fire presence and absence were derived from burn scar data for 2014-2023 from GISTDA [1]. GISTDA produces burn scar data by calculating the difference between pre- and post-fire Normalized Burn Ratio (NBR) is calculated from Sentinel-2 imagery to produce delta NBR (dNBR); dNBR is then manually thresholded on a regional basis using its frequency distribution as a guide for determining the localized cutoffs for burnt areas [1,38,76]. GISTDA reports that this methodology has an overall accuracy around 66.8% [1].

This approach is preferable to using MODIS or VIIRS-derived FIRMS active fire hotspots [60] as reference data for fire occurrence in Thailand for multiple reasons [77]. The MODIS and VIIRS sensors detect fires through temperature anomalies in their thermal bands, providing full global coverage once daily for MODIS and twice daily for VIIRS [5,78,79,80]. MODIS has a spatial resolution of 1000m and VIIRS has a spatial resolution of 375m [5,78]. MODIS-derived hotspots have low error of commission in Thailand [81] and VIIRS-derived hotspots can detect fires as small as 5m² both globally [82] and in Thailand [83]. However, these data lack the spatial resolution necessary to pinpoint precise locations for the small transient fires characteristic to Southeast Asia [77]. GISTDA burn scars in Thailand have a median size of 0.006 km2 and a mean size of 0.15 km2, compared to the 0.14-1 km2 pixel size of the FIRMS data. FIRMS point locations are generated at the centroids of the MODIS or VIIRS pixels, so fires smaller than a pixel will have imprecise locations. Additionally, fires must produce enough heat to be detected during the satellite overpass [5,78]. Sensors sometimes do not register agricultural fires on small fields or understory fires under dense forest canopies, even if they burn during multiple satellite overpasses [14]. Further, landowners and officials are aware of satellite orbital patterns and will sometimes time their intentional burns to be between overpass times in order to avoid detection [14].

Reference points for fire presence and absence were randomly generated within historical burn scar polygons from 2016-2023 obtained from GISTDA (Supplement Figure S1). For each year, 600 total reference points were generated (300 in areas that burned that year and 300 in areas that did not burn that year) yielding a total of 4800 reference points [84,85,86]. Of these reference points, 80% were used to train the model and 20% were used to validate the model for 2016-2023 [87]. An additional 600 reference points were generated solely for validating the final 2024 probability map (300 in areas that burned in 2024 and 300 in areas that did not burn in 2024).

2.3.2. Predictor Variable Data

We used environmental and socioeconomic variables with known relationships to fire probability in Thailand as predictor variables based on similar fire probability mapping analyses from scientific literature in peninsular Southeast Asia and operational methods in the Thai government [1,10,38,39,40,41,42,43,44,45,46,47,48,49,50,51]. These include variables representing topography [10,42,88,89], fuels [12,24,89,90,91,92,93,94,95], potential fire behavior [91,93], forest type [1,3,10,12,26,28,90,92,96,97,98], vegetation characteristics [3,12,26,28,90,92,96,97,98], climate [1,2,12,24,89,95,98], water availability [1,10,39,99], crop type [1,13,18,20,21,22], recent burn history [1,10,12,18,25,30,100], and human influence and accessibility [10,12,98] (Table 1). These variables both directly and indirectly influence ignition likelihood and fire behavior, and many of them exhibit complex interactions with each other across multiple temporal and spatial scales [18,26,88,92,96,97,101,102,103,104].

2.4. Data Preprocessing

Full descriptions of preprocessing steps and example maps for each predictor layer can be found in Supplement Table S3 and Supplement Figure S8a-j. Publicly available global data layers were used to represent predictors for which we did not have local data sets. All predictor variable layers were resampled to a 300 m resolution using either mean or minimum aggregating functions, as appropriate. For data sets that had spatial resolutions lower than 300 m (pixel size larger than 300 m), the pixels were subdivided to 300 m and smoothed with a focal mean function.

All variables with multiple time steps available were composited to produce operational layers comprising data available before the start of each fire season. The exact compositing methods varied based on the known seasonality of the variables and the data type, quality, and availability. For example, climate variables were derived from composites of the 5 pre-fire season months (August-December) [1,2,3,4], forest canopy and change variables were derived from annual values from the two pre-fire season years, optical imagery indices were derived from composites of the pre-fire season year, and seasonal SAR imagery differences were derived from composites of the 4 months in the peak leaf-on season (July-October) and 4 months in peak leaf-off season (January-April) [64,65,66,67]. For data sets that had temporal resolutions lower than 1 year (intervals greater than 1 year), annual layers for years without data were created using the most recent values. Predictor variable values were extracted to the reference points from the corresponding year. For predictors for which we had no historical data, the values from the single-date static layers were extracted to all points.

In order to ensure the reliability of variable importance assessments, we ensured that all variables were continuous and scaled to the same numerical range. The Gini impurity index favors variables that provide more potential splits during tree construction, including variables with more unique values and larger numerical ranges [123,124]. Thus, it often assigns higher importance to continuous rather than categorical variables, and higher importance to variables with higher numerical values rather than lower numerical values [123,124]. Additionally, if some rare classes have very few representative reference points, the model may not be able to identify their predictive relationship with fire, even if a relationship does exist [125]. Categorical variables were converted to continuous values by calculating the distance to each class, alleviating the biases posed by categorical variables and sampling bias. All predictor variable values were also normalized to a zero to one scale using a simple min-max normalization, alleviating the biases posed by differing numerical ranges.

2.5. Model Development and Refinement

2.5.1. Multicollinearity

We assessed multicollinearity between variables using correlation coefficients and the Variance Inflation Factor (VIF) [126,127] (Supplement Figures S2-S4). While random forest models have been shown to maintain predictive performance despite multicollinearity in predictor variables [73,74], multicollinearity can confound variable importance rankings in random forest models [125,128,129]. Two pairs of variables had high correlation, namely elevation and maximum temperature and EVI and NDWI, however we kept both pairs as they independently influence fire behavior and their VIF fell below the generally accepted threshold [127]. Ecological data often exhibit inherent multicollinearity, and it can be appropriate to retain highly correlated variables in random forest models when the goal of the study includes identifying multiple important variables [125].

2.5.2. Variable Importance

Variable importance was calculated for all predictors using the Gini impurity index. Despite its potential biases, Gini impurity is the most accessible empirical variable importance metric on which to base variable selection for an operational fire probability model. The Gini index measures how often a randomly chosen data point would be incorrectly classified if it were given a random class label based on the distribution of classes in a given subset of data points [130,131]. The variable importance metric can be defined as how much each variable contributes to the model’s ability to distinguish between classes.

We normalized variable importances to sum to a 100 for easy interpretation. The model was then refined by removing variables individually in order of importance to the model until further removals resulted in a distinct drop in 2016-2023 AUC below 0.84 (Figure 4). Eleven variables were removed during model refinement, leaving 31 variables in the final version of the model. A full chart of variable importances before model refinement can be found in Supplement Figure S5.

2.5.3. Accuracy Assessment

We used Area Under the Curve (AUC), which is calculated for the Receiver Operating Characteristic (ROC) curve, to assess the predictive power of the model [46,47,48,49,51,132,133,134,135]. We calculated AUC for the two separate testing sets: 20% of the 4800 reference points from 2016-2023 set aside for validation, and the 600 reference points from 2024 generated solely for validation. AUC quantifies a model’s ability to correctly classify the validation data compared to a random classification; values between 0.8 and 0.9 are excellent and values between 0.9 and 1.0 are outstanding [136,137,138].

2.5.4. Sensitivity Analysis

We also conducted a sensitivity analysis on the model’s response to the removal of individual variables. We trained and validated 44 versions of the random forest model, with the initial run including all variables and each subsequent run removing only a single variable from the set. We calculated the mean and median of each variable’s scaled importance values across all runs and plotted the difference in 2016-2023 AUC (dAUC) between each run and the initial run.

3. Results

3.1. Variable Importance and Model Sensitivity

Two distinct groups emerged within our variable importance rankings: a group of low importance variables, and a group of high importance variables with nearly equal importance values. There is a clear difference in importance values between the top and bottom groups and a distinct drop in 2016-2023 AUC once variables from the high importance group are removed during model refinement (Figure 4). Before model refinement, the most important variables each contributed between 2.4-3.1% of the total impurity reduction across all splits within all trees (Supplement Figure S5). These included certain forest and crop types, vegetation characteristics, topography, climate, human influence and accessibility, water availability, and recent burn history (Figure 5).

In the sensitivity analysis, seasonal difference in SAR backscatter, PDSI, aspect, distance to bamboo forest, and EVI had the highest mean and median importance values across all runs, while population density, distance to Royal Forest Department (RFD) and Department of National Parks, Wildlife, and Plant Conservation (DNP) protected areas, grass height, litter depth, and canopy height change had the lowest importance values (Figure 6). Removal of distance to hill evergreen forest, soil moisture, canopy height change, and slope increased the 2016-2023 AUC, while the removal of distance to burns 2 years prior, distance to burns 1 year prior, and canopy cover decreased the 2016-2023 AUC (Figure 7).

3.2. Fire Probability Map

The 2024 Fire Probability in Northern Thailand map (Figure 8) is available online as an interactive application in Google Earth Engine.

The overall AUC of our fire probability model is 0.841 for 2016-2023 and 0.848 for 2024, which demonstrates a reasonably high ability to discriminate between fire presence and absence [136,137,138]. The ROC curves show that as the classification threshold is gradually lowered, the true positive rate increases rapidly but the false positive rate increases more slowly (Supplement Figure S6). This indicates that the model can distinguish well between areas that have burned and have not burned using the selected predictor variables in the area and time period of interest. The accuracy of the model for the 2024 prediction can be calculated once 2024 burn scar data are made available through official channels.

4. Discussion

4.1. Evaluation of Model Results

Most fires in Thailand are anthropogenic and ignited for land management purposes in agricultural production or non-timber forest product collection [1,12,13,14,15,16,17,18]. Our 2024 fire probability map aligns well with these known patterns, indicating that the highest probability of fire is in agricultural and forested areas near human settlements, accessible to roads and at mid elevations. Empirically, our model also performed well, with an AUC of 0.841 for the 2016-2023 testing points and an AUC of 0.848 for the 2024 testing points [136,137,138].

Overall, despite two major differences in methods, our seasonal spatial patterns align with other machine learning based fire probability approaches in Thailand and SE Asia. First, while other studies used a small number of predictor variables based on theory [46,47,48,49,50,51], we began with a large number of predictor variables and iteratively removed empirically less important variables. We found that certain forest and crop types, vegetation characteristics (seasonality, health, density), topography (elevation, aspect, slope), climate (precipitation, temperature, VPD, soil moisture, drought indicators), human influence and accessibility (roads, settlements, special management designations), water availability, and recent burn history are the strongest predictors of fire probability in northern Thailand. These variables were all of near equal importance in our model (Figure 5, Supplement Figure S6), and our sensitivity analysis confirmed the decision to retain these variables (Figure 6 and Figure 7). This is in contrast with similar studies that found only a few variables of high importance, with the highest often being distance to roads and human settlements [46,47,48,49,51].

Differences in variable importance may be attributed to differences in model structure, input data quality, and variable importance metrics, complicating direct comparisons. Specifically, other studies use fewer variables with minimal multicollinearity (e.g., [46]), or different variable importance metrics (e.g., Relief-F Average Merit [48], Correlation Attribute Evaluation [49], Pearson Correlation Coefficient [47,51], or Cramer’s V Coefficient [45]). Some of our observed patterns in variable importance may also stem from the statistical properties of the Gini Index, namely its sensitivity to high numbers of variables [125], sampling bias in variables [125], multicollinearity among variables [125,128,129], and differing unique value counts across variables [123,124,128]. This may have contributed to the low importance we observed for population density, distance to protected areas, fuel characteristics, predicted fire behavior, and canopy change, along with the discrepancies between dAUCs and Gini Index rankings in our sensitivity analysis.

Further, many of these metrics, including the Gini Index, cannot capture spatial autocorrelation [130,139], which is important for many geospatial phenomena including fire. Future research should evaluate variable importance metrics in the context of fire probability mapping in Thailand in order to understand the practical implications and tradeoffs of using different metrics to build machine learning models for management applications^,,,,. We also suggest incorporating forecasts for fire season weather conditions; these weather conditions influence fire behavior in real time and have a different effect than pre-fire season climate which largely influences fuel conditions. Including crop rotation and shifting cultivation as predictor variables may also improve the model; agriculture is particularly transient in nature, as different crop types are often rotated seasonally on a single plot of land and forests are frequently cleared to be cultivated or grazed for only a few years. The model can also be applied with future climate projections as the input climate predictor variables to see how different climate change scenarios will alter fire probability. An important component of this is also calculating the uncertainty associated with each set of conditions.

The second major methodological difference for our approach is that we paired annual fire occurrences with annual predictor variables in order to capture temporal variation, while other studies use cumulative fire occurrence with temporally averaged predictor conditions [46,47,48,49,50,51]. This allows us to evaluate year to year shifts following both monotonic (deforestation, climate change) and cyclical (ENSO) trends. When interannual variations in predictor conditions are aggregated for the study period, their influence on fire probability is also aggregated. Thus, cyclical patterns during the study period may be discounted. Further, with each new fire season, the model can be retrained and rerefined with the addition of the most recent fire occurrence and fire predictor data to enhance its performance.

However, this approach has important implications for both the model training data selection and calculating accuracy metrics. Considering temporal variation allowed us to draw a more representative sample for fire presence and absence. Other approaches define fire absences as locations that never burned within a multi-year study period [46,47,48,49,50,51], but our model also considers areas that may have burned in one year in the study period but not in other years as fire absences for those years in which the area did not burn (Supplement Figure S7). This allowed us to draw a more representative sample for fire absences, as there are fundamental ecological and socioeconomic differences between areas that have never burned in the study period and areas that have burned, but not during the specific year of interest. Sampling fire absences only from areas that never burned creates a biased sample that provides an extreme basis for comparison with fire presences.

The increased temporal resolution likely contributed to the slightly lower AUC compared to the ≥0.9 values reported by other random forest models [46,47,48,49,51]. Higher temporal resolution preserves cyclical variability—e.g., in climate—that would otherwise be smoothed by multi-year averages. Including areas that burned during the study period but not in a specific year better represents the fire regime in the model but slightly reduces predictive power. In contrast, other models often use a biased subset of fire absences and smooth inter-annual variability, potentially exaggerating predictor-response relationships and inflating AUCs. While most models capture only spatial variation, ours accounts for both spatial and temporal variability, introducing greater complexity and noise due to the more representative sampling and added temporal dimension. These differences suggest future work should examine how model performance varies with the selection of training and validation years to optimize prediction accuracy while minimizing data requirements (Supplement Table S4).

4.2. Feedback from Stakeholders

Stakeholders provided feedback on a preliminary version of the fire probability mapping approach at a Regional Consultation in Chiang Mai Province and National Consultation in Bangkok. The proposed methods were well received, especially by technical geospatial experts who are already familiar with the advantages of machine learning and cloud computing and are interested in expanding their departments’ use of these tools. However, many participants also recognized the organizational and technical obstacles that would need to be addressed in order to implement these methods. This would include widespread capacity building and data collection, as well as changes in organizational structure to integrate automated tools with manual methodologies.

Stakeholders also expressed the need to identify the intended uses and users of the fire probability map, from high level officials and local communities. In its current form, the map can provide a broad overview of expected fire patterns, but it should be further calibrated for community use. Community-level data should be incorporated to refine the analysis, since local communities are one of the primary actors in preventing and fighting fires [14]. Local communities would need much greater spatial and temporal precision in the map for it to be useful for on-the-ground activities, which can be achieved if such products with the desired spatial and temporal resolution are incorporated into the model in lieu of the global data sets currently used as placeholders [14].

These concerns about data quality were also reflected in the final model’s accuracy metrics. The reference data points were produced from the burn scar polygons provided by GISTDA, which were themselves generated using an automated methodology. These data are likely less accurate than data produced by manual image interpretation or field surveys. Similarly, the fuel, potential fire behavior, climate, and canopy data layers were derived from global data sets, which were generated using generalized models at the global or continental scales [105,109,110,113]. These data are likely less precise than regional or national scale data that are generated using locally calibrated and validated models.

These concerns can be addressed with improved data from local experts. The current model uses coarse, global public datasets for many predictors like climate, weather, and fuels, producing a fire probability map with lower resolution than the scale of fires in Thailand. For variables like LULC and infrastructure, only static, single-date datasets were available, which miss rapid interannual changes that may affect fire probability. Replacing them with high-resolution, historical local data would improve model performance and operational value. Some datasets, like climate and weather, are regularly produced by the government but remain hard to access, while others, like fuels and corresponding fire behavior, do not yet exist and would require extensive fieldwork and modeling to produce. We also recommend generating reference data from manually delineated burn scars to improve training and testing accuracy.

Further, developing separate fire probability models for each LULC type could reveal both broad differences between classes and finer-scale variation within them. Fire regimes, their drivers, and human management vary widely among LULC classes. Since random forest models and the Gini impurity index treat categorical and continuous variables differently, including LULC alongside continuous predictors complicates the interpretation of variable importance. Therefore, building and comparing separate models for key LULC types, such as specific crops or forest types, is advisable.

Finally, local users would benefit from higher temporal resolution in fire probability maps. Shifting from seasonal to daily, weekly, or monthly predictions would improve prevention, mitigation, and response during fire season. Weather, a key driver of fire behavior, changes frequently across these timescales. More frequent predictions are possible if matching temporal data are available for key variables. This includes date-labeled fire occurrences and short-term weather forecasts aligned with the desired prediction intervals.

5. Conclusions

We developed an annual random forest machine learning model approach to mapping fire probability in northern Thailand that produces a straightforward fire probability metric. The model yielded a high AUC of 0.841 for 2016-2023 and 0.848 for 2024, and indicated the highest probability of fire is in easily accessed agricultural and forested areas adjacent to human settlements. Variables with high predictive power in the model included certain forest and crop types, vegetation characteristics, topography, climate, human influence and accessibility, water availability, and recent burn history. The model aligns with both previous maps of fire probability and known mechanistic drivers of fire in Thailand.

Key process improvements include automated variable selection, disaggregated historical data, and representative sample design. Empirically selecting the most influential fire predictor variables from a comprehensive set of predictor variables, rather than pre-selecting variables, allows flexibility and opportunities for localization. Due to the modularity and automation employed by the model, it can also be modified using higher quality local data in lieu of global datasets for localization. It can also be adapted to data scarce regions, using fewer years of fire occurrences or predictor variables until more data can be made available. Disaggregating historical data to account for year to year variability in fire predictors allows the model to be sensitive to cyclical variables. This allowed us to use a more representative sample for fire absences, including areas that did not burn in a given year rather than only areas that never burned.

With machine learning and other forms of artificial intelligence becoming standard tools in the field of fire probability mapping worldwide [140,141,142,143,144]. Thailand should leverage these tools for their own fire management activities. Our approach provides the groundwork for an easy to implement approach that predicts fire probability annually and can be adapted based on locally available data.