Enhancing forest fire susceptibility mapping in Xichang City, China using DBSCAN-based non-fire point selection integrated with deep neural network

Forest fire susceptibility mapping plays a crucial role in forest management and disaster prevention. However, existing research often neglects the selection of non-fire data during model construction, resulting in limited prediction accuracy. To address this issue, we propose an innovative DBSCAN-DNN model that optimizes non-fire data selection to enhance mapping precision. Using VIIRS and GLC_FCS30D datasets, we created a spatial database for Xichang’s dry seasons from 2012 to 2022, incorporating topography, meteorology, vegetation, and human activities. Based on this, we employed the DBSCAN algorithm to cluster the fire points and accurately delineated the affected areas. Subsequently, we selected non-fire samples from outside these regions for training the DNN model. Through comparative experiments, we found that the DBSCAN-DNN model exhibited excellent performance in predicting forest fire susceptibility in Xichang City, with an AUC value of 0.925 and significant improvements in accuracy (0.834), precision (0.800), recall (0.891), F1-score (0.843), and Kappa coefficient (0.669). Additionally, we conducted a SHAP analysis to delve into the contributions and interactions of various factors influencing fire susceptibility. This finding offers valuable insights for selecting non-fire sample data in the forest fire susceptibility mapping model.