Machine Learning Framework Maps Wildfire Susceptibility Across Central India

A regional wildfire assessment framework developed for Madhya Pradesh, India, has demonstrated that gradient-boosting tree algorithms and stacked ensembles deliver superior predictive discrimination and computational efficiency compared to complex deep-learning sequence architectures. The investigation established a systematic spatial modeling structure across approximately 308,000 square kilometers of vegetated landscape, evaluating predictive capability through conventional random holdout partitions and rigorous spatial block cross-validation. Researchers found that while deep neural networks achieved high sensitivity to recorded fire events, tree-based models maintained a far more balanced distribution of precision and recall. Across independent testing, an ensemble integrating extreme gradient boosting, light gradient boosting, and standard gradient boosting achieved the highest overall discrimination, demonstrating that architectural complexity does not inherently translate into superior geographic mapping performance.

Madhya Pradesh, situated in central India and widely recognized as the geographic heart of the country, encompasses about 94,689 square kilometers of forest, representing roughly 12.4 percent of India’s total forest cover. The state is characterized by low-lying northern plains, central plateaus, and southern hill ranges reaching elevations from 72 meters to more than 1,300 meters above sea level, drained by major river systems including the Narmada, Tapti, Mahanadi, and Wainganga. The regional climate features dry, hot summers where temperatures frequently exceed 40 degrees Celsius, accompanied by an intense spatial precipitation gradient ranging from 700 millimeters annually in the west to 1,800 millimeters in the east. Driven by prolonged dry periods, severe pre-monsoon heat waves, and extensive human-forest interfaces, forest fire occurrences across India have escalated nearly tenfold over the past two decades, with approximately 15.8 percent of Madhya Pradesh’s forested territory categorized as highly fire-prone.

To capture the environmental and human mechanisms driving fire occurrence, the research team compiled 15 conditioning factors spanning topography, climate, surface energy, biophysics, hydrology, and anthropogenic accessibility. Topographic metrics including slope, aspect, elevation, and topographic wetness index were derived from Shuttle Radar Topography Mission digital elevation models at 30-meter resolution. Atmospheric and energy dynamics were represented through precipitation and solar radiation data from NASA POWER, actual evapotranspiration from TerraClimate, and land surface temperature from the Moderate Resolution Imaging Spectroradiometer. Biophysical and hydrological conditions were captured using MODIS normalized difference vegetation index, vegetation health index, Soil Moisture Active-Passive satellite observations, and HydroSHEDS river networks. Human pressure was characterized through MODIS land cover classifications alongside distance-to-road and distance-to-settlement layers generated from OpenStreetMap. All environmental predictors were reprojected to UTM Zone 44N and aligned to a standardized 1-kilometer modeling grid.

The binary modeling dataset was developed by pairing high-confidence active fire observations with vegetation-constrained pseudo-absences. Fire locations detected between 2001 and 2023 were extracted from MODIS active-fire products, retaining only detections with satellite confidence ratings of at least 80 percent situated within vegetated areas defined by annual land-cover masks. To prevent overrepresentation from repeated satellite passes over ongoing blazes, points were grouped using Density-Based Spatial Clustering of Applications with Noise, applying a 1-kilometer spatial threshold and a 24-hour temporal window to isolate 6,839 independent fire locations. Pseudo-absences were sampled exclusively from burnable vegetated pixels located at least 1 kilometer away from recorded fire coordinates, preventing the artificial inflation of model accuracy through non-burnable water or urban surfaces. The final balanced dataset comprised 5,863 fire observations and 5,863 pseudo-absences, yielding 11,726 total samples.

Multicollinearity diagnostics were executed prior to model training to evaluate linear and monotonic dependencies across the 15 conditioning layers. Variance Inflation Factor assessments demonstrated that actual evapotranspiration exhibited the highest redundancy at 9.33, closely followed by soil moisture at 8.34, though both values remained below the traditional collinearity cutoff threshold of 10. Precipitation ranked third with a factor of 2.90, while terrain aspect approached 1.0, reflecting virtually complete independence. However, Spearman rank correlation analysis identified an exceptionally strong monotonic association between actual evapotranspiration and soil moisture that exceeded a predefined threshold of 0.90. In a controlled decorrelation sensitivity experiment, removing soil moisture reduced the predictor set to 14 variables and caused only a slight reduction of 0.010 in the stacked ensemble’s held-out receiver operating characteristic area under the curve, confirming that the framework’s predictive power was not artificially inflated by predictor redundancy.

The study implemented and compared six predictive models, comprising XGBoost, LightGBM, Gradient Boosting, a stacked ensemble combining all three tree algorithms via logistic regression, a Temporal Convolutional Network, and a hybrid architecture merging temporal convolutions with bidirectional long short-term memory units. For the deep-learning models, standardized environmental predictors were arranged into a fixed 15-by-1 feature sequence. The authors explicitly noted that this architectural configuration represents ordered feature-sequence learning across environmental attributes rather than temporal forecasting across chronological time series. The tree-based algorithms underwent randomized hyperparameter search across 16 configurations with five-fold cross-validation, while the neural networks were optimized using the Adam algorithm with validation monitoring, learning rate reductions on plateaus, and early stopping over a maximum of 180 epochs to prevent over-specialization to training samples.

Evaluation on the independent 30 percent held-out test partition comprising 3,518 observations revealed marked behavioral differences across model families. The stacked ensemble achieved the highest observed discrimination with an area under the curve of 0.800 and a 95 percent bootstrap confidence interval of 0.785 to 0.814, closely followed by XGBoost at 0.795, LightGBM at 0.790, and Gradient Boosting at 0.775. The standalone Temporal Convolutional Network attained an area under the curve of 0.773, whereas the hybrid network dropped significantly to 0.674, with a non-overlapping confidence interval of 0.657 to 0.690. The stacked ensemble also generated the lowest probabilistic log loss at 0.542 and the highest precision at 0.709, correctly classifying 1,206 non-fire and 1,349 fire instances at a 0.5 probability threshold while producing 553 false positives and 410 false negatives.

A critical divergence emerged in how the deep-learning architectures handled classification trade-offs on the holdout dataset. Both the temporal convolutional network and the hybrid bidirectional model exhibited high sensitivity to fire events, posting recall values of 0.836 and 0.837, respectively, which noticeably exceeded the 0.767 recall achieved by the stacked ensemble and the 0.770 recall of XGBoost. However, this sensitivity came at the expense of substantial false-positive rates, causing precision to decline to 0.658 for the convolutional network and 0.589 for the hybrid recurrent model, which misclassified 1,026 non-fire sites as fires. Consequently, although the convolutional network matched the stacked ensemble’s overall F1-score of 0.737, it did so through an asymmetric classification behavior that inflated fire predictions across non-fire locations, while the tree-based ensembles maintained balanced accuracy and superior probability calibration.

Model stability was corroborated through 10 repeated random 70:30 splits, where XGBoost and the stacked ensemble produced nearly identical mean areas under the curve of 0.8000 and 0.7997, accompanied by narrow standard deviations under 0.007. However, when the research team applied a five-fold spatial block cross-validation using 50-by-50-kilometer geographic blocks, performance dropped systematically across all algorithms. Mean spatial areas under the curve declined to 0.752 for XGBoost, 0.751 for the stacked ensemble, 0.748 for LightGBM, and 0.735 for Gradient Boosting, representing an average decrease of 0.040 to 0.049 units. Furthermore, fold-to-fold standard deviations expanded fourfold to approximately 0.030. This performance drop highlights that conventional random data partitioning overestimates predictive capability due to spatial proximity between training and testing samples, proving that spatial block validation is essential for realistic transferability assessments.

Model explainability was examined using TreeExplainer Shapley Additive Explanations for individual tree models, permutation importance for the stacked ensemble, and Integrated Gradients for neural architectures. Across all methods, normalized difference vegetation index emerged as the dominant predictor of wildfire occurrence, followed consistently by precipitation, solar radiation, actual evapotranspiration, soil moisture, and land surface temperature, while terrain aspect and topographic wetness contributed minimal predictive power. Permuting the vegetation index within the stacked ensemble caused the sharpest deterioration in discrimination. The authors emphasized that these metrics quantify predictive feature attribution within complex models rather than independent causal mechanisms, noting that correlated hydroclimatic variables naturally share attribution across algorithms and do not represent direct field-scale intervention outcomes.

The study also assessed how pseudo-absence definitions influenced apparent model success. Varying the minimum exclusion distance separating pseudo-absences from detected fires revealed that expanding the exclusion buffer from 0.5 to 1.0 and 2.0 kilometers caused the stacked ensemble’s area under the curve to rise sharply from 0.764 to 0.800 and 0.867. However, this buffer expansion reduced the available balanced sample size by roughly 30 percent, from 13,068 to 9,094 observations. The investigators noted that higher discrimination at larger distances reflects artificial separation of fire and non-fire classes in geographic and environmental space rather than genuine algorithmic enhancement, justifying the 1-kilometer distance as an optimal compromise between sample availability and class integrity.

Continuous probability surfaces were converted into four susceptibility zones—Very Low to Low, Low to Moderate, Moderate to High, and High to Very High—using quartile thresholds derived strictly from training probabilities. Independent validation on holdout data confirmed a steep monotonic gradient, with observed fire rates rising from 11.4 percent in the lowest susceptibility class to 42.3 percent, 66.1 percent, and 82.2 percent in the highest tier. When mapped across the state, the stacked ensemble allocated 36.4 percent of the territory to Very Low to Low, 37.5 percent to Low to Moderate, 19.3 percent to Moderate to High, and 6.8 percent to High to Very High susceptibility. In contrast, the hybrid recurrent neural network expanded the top category to 14.2 percent, mirroring its high false-positive bias. Overall spatial agreement was strong across central, southern, and eastern forest belts, identifying consensus areas where fire management and surveillance can be prioritized.

Computational measurements recorded within a two-core CPU environment demonstrated marked efficiency differences among workflows. Training required only 1.6 seconds for LightGBM, 4.2 seconds for XGBoost, and 7.5 seconds for Gradient Boosting, whereas the stacked ensemble completed in 64.7 seconds, including out-of-fold generation and meta-learner fitting. Conversely, neural network training required 153.0 seconds for the hybrid model and 283.8 seconds for the convolutional network. The authors cautioned that their framework models static baseline susceptibility rather than real-time fire detection, rate of spread, or future climate projections, concluding that carefully validated, computationally lightweight tree ensembles provide an optimal and transparent foundation for regional forest fire prioritization.

Subject of Research: Regional wildfire susceptibility modeling using machine learning and spatial cross-validation

Article Title: A spatially validated explainable framework for large-scale wildfire susceptibility mapping using deep sequence learning and gradient-boosting ensembles

Article References: Zerouali, B., Mishra, M., Alturki, A. A., Kavhiza, N. J., Guria, R., Paul, S., Bhattacharyya, D., da Silva, R. M., Santos, C. A. G., & Youssef, Y. M. (2026). A spatially validated explainable framework for large-scale wildfire susceptibility mapping using deep sequence learning and gradient-boosting ensembles. Frontiers in Forests and Global Change, 9, 1970558. https://doi.org/10.3389/ffgc.2026.1970558

Image Credits: AI Generated

DOI: 10.3389/ffgc.2026.1970558

Keywords: wildfire susceptibility, machine learning, gradient boosting, spatial cross-validation, remote sensing, Madhya Pradesh, explainable AI, environmental modeling, spatially, validated, explainable, framework