Advancing wildfire prediction in Nepal using machine learning algorithms

Wildfires are increasingly threatening Nepal, particularly during the dry pre-monsoon months (March-May), leading to severe ecological impacts and disruptions to local communities. To improve wildfire prediction and preparedness, this study evaluated four advanced machine learning algorithms-Random Forest, Radial Basis Function Neural Network, Artificial Neural Network, and Support Vector Machine-using comprehensive dataset (2001-2023) of meteorological, topographical, anthropogenic, locational, and vegetation variables. The Random Forest (RF) model outperformed others, achieving the highest accuracy (88.6%) and predictive reliability (AUC: 0.96). Notably, vapor pressure deficit emerged as the strongest predictor, contrasting previous studies where precipitation was often considered dominant. Utilizing the robust RF model, a high resolution (1-km) wildfire risk map identified 11.1% of Nepal, encompassing 12 districts and 48 municipalities primarily in the southwestern region, as very high-risk areas. By integrating daily meteorological data into wildfire predictions, this research provides an innovative framework that enhances risk management strategies, offering actionable insights for decision-makers and supporting resilience-building efforts in fire prone regions.