Spatial prediction of the number of wildfire hotspots using negative binomial conditional autoregressive random forest

BACKGROUND AND OBJECTIVES: Predicting the spatial distribution of wildfire hotspots is crucial for forest fire mitigation, particularly in high-risk regions such as Kalimantan Island, Indonesia. These hotspots are locations with thermal anomalies above normal background temperature. An integrated framework is proposed: a negative binomial conditional autoregressive random forest. The approach is designed to capture nonlinearity in meteorological drivers, over-dispersion in the number of hotspots, and spatial dependence among neighboring grid cells. The aim of this study is to predict the number of hotspots and to identify the meteorological factors that most strongly influence the emergence of hotspots.METHODS: The number of hotspots in Kalimantan for September 2024 were analyzed on a grid. Predictors included air temperature, precipitation, relative humidity, wind speed, and dry days. These variables were aggregated into spatial grids with a resolution of 0.25 degree by 0.25 degree. In a simulation study, levels of spatial autocorrelation, alternative spatial weight matrices, ratios of predictor variance to spatial residual variance, and sample sizes were varied. Model performance was evaluated using root mean square error and mean absolute error.FINDINGS: The proposed model consistently outperformed the conventional conditional autoregressive model in both simulated and empirical datasets. Gains were largest when predictor variance exceeded spatial residual variance. Across most scenarios, reductions in root mean square error by 24 percent and in mean absolute error by 33.3 percent were observed. Results were found to be robust to alternative spatial weight matrices. In the empirical application, the model accurately predicted the spatial distribution of hotspots. Overestimation occurred in approximately 21.43 percent of the grid cells of the observed. Air temperature and relative humidity were identified as the most influential predictors, but their relative importance varied across spatial clusters. Meteorological effects were weaker in wet, humid clusters and stronger in hot, dry clusters.CONCLUSION: Practical accuracy improvements for spatial prediction of the number of hotspots are generated by the integrated framework. Incorporation into operational early-warning systems and decision support for forest and land fire prevention in fire-prone tropical regions is recommended.