Hybrid conditional autoregressive generalized random forest for overdispersed spatial counts: Evidence from Kalimantan wildfire hotspots

This paper proposes a negative binomial conditional autoregressive generalized random forest model for overdispersed spatial count data. The framework combines a flexible nonparametric mean function estimated via generalized random forests with conditional autoregressive spatial random effects under a negative binomial likelihood. We evaluate the proposed method through a simulation study and an empirical application to satellitederived wildfire hotspot counts in Kalimantan, Indonesia (September 2024), using rainfall, air temperature, relative humidity, wind speed, and the number of rainless days as covariates. Across simulation scenarios, negative binomial conditional autoregressive generalized random forest yields lower prediction errors than the benchmark negative binomial conditional autoregressive random forest. In the Kalimantan application, the model achieves a root mean squared error of 3.64 and a mean absolute error of 2.11 hotspots per grid cell. Variable importance analysis indicates that air temperature is the most influential predictor of the spatial distribution of hotspots.