Forest Fire Susceptibility Mapping Using Deep Neural Network

Abstract

Forest fires are an increasing environmental problem in Nepal, threatening biodiversity, ecosystem stability, and the livelihoods of forest-dependent communities. In recent years, both the frequency and intensity of fires have increased, highlighting the need for accurate forest fire susceptibility mapping to support effective management and risk reduction. This study develops a forest fire susceptibility model for Kailali District using a deep learning approach, integrating twelve explanatory variables related to topography, climate, vegetation, and human influence. Historical fire data from NASA’s VIIRS archive (2012–2024) were used for training and validation, with trend analysis indicating peak fire events in 2016 and the lowest in 2020. The dataset was split into 70% training and 30% testing, and model performance was evaluated using accuracy, precision, recall, and F1-score, while the trained model was also applied to Dang District. The susceptibility map generated using the DNN model was classified into five risk zones from very low to very high, achieving accuracy, precision, recall, and F1-score values of 0.93, 0.92, 0.93, and 0.93, respectively. Overlay analysis of forest fire susceptibility map and historical fire occurrence showed that about 91% of observed fire events in Kailali and 88% in Dang District occurred within susceptible zones. This study demonstrates that integrating geospatial data with deep learning can effectively improve forest fire risk assessment in Nepal. The resulting susceptibility maps provide useful information for early warning systems, disaster preparedness, and sustainable forest management.