Forecasting Locations of Forest Fires in Indonesia Through Nonparametric Predictive Inference with Parametric Copula: A Case Study

Abstract

Wildfires caused major damage and incurred high restoration costs. Despite numerous predictive studies in this field, wildfire management still had uncertainties. The machine learning technique was popular on this topic, but it portrayed gaps of non-generalisable and inaccuracy possibilities. This study intended to apply nonparametric predictive inference (NPI) with a parametric copula to predict the next wildfire location using the coordinate parameters. The NPI quantifies the uncertainties via imprecise probabilities, (𝑃, 𝑃) , while the copula integration considers the spatial correlation by modelling the dependence structure between the past coordinates in predicting the next location. Unlike other methods, the NPI generates a set of bounded probabilities that provide confidence in the prediction result. This paper applied the proposed method to the Moderate Resolution Imaging Spectroradiometer satellite dataset for Indonesia (2020). Several wildfire hotspots in Sumatra and Kalimantan archipelago were focused on this study. It was evaluated via the differences (𝑑̅) within the (𝑃, 𝑃) and showcased low values (𝑑̅< 0.001). The results show that NPI with parametric copula was highly accurate for both archipelagoes, highlighting its generalisability specifically for Indonesia. Each wildfire hotspot had a different optimal copula to predict the best future hotspot. Clayton and Gumbel copulae were the best to be integrated with NPI to predict the next wildfire location in Sumatra while Normal and Gumbel copulae for Kalimantan locations. In conclusion, the NPI is considered a reliable alternative for wildfire location prediction.