Forest fires critically threaten biodiversity and ecological stability, particularly in vulnerable regions in the dry and hot region. Despite the growing application of machine learning (ML) in fire susceptibility mapping, research gaps persist in biodiversity-rich regions and the integration of long-term climate data. This study addresses these gaps by developing forest fire susceptibility (FFS) maps using Random Forest (RF) and Classification and Regression Tree (CART) models, integrated with environmental variables derived from Google Earth Engine (GEE). The objectives were to (1) analyze spatiotemporal fire patterns (2001–2024) using MODIS FIRMS data, (2) evaluate topographic, climatic, and vegetation variables, and (3) compare model performance for fire risk zoning. Fire occurrence data and 14 environmental predictors (e.g., elevation, NDVI, precipitation, LST) were analyzed. The Boruta algorithm identified elevation, SAVI, NDVI, and precipitation as key drivers. The RF model demonstrated superior accuracy (77.54%, AUC: 0.802) compared to CART (76.08%, AUC: 0.706), with spatial mapping revealing divergent risk patterns: RF classified 40.21% and 47.14% of the reserve as moderate and high-risk zones, whereas CART polarized 91% of the area into low (47.9%) and very high (43.32%) risk categories. The RF model’s nuanced classification underscores its robustness in capturing environmental interactions, making it ideal for targeted fire management. This study provides a scalable framework for integrating ML and remote sensing in fire risk assessment, aiding policymakers in prioritizing mitigation efforts in biodiversity hotspots. By addressing the gap in region-specific ML applications and emphasizing climate-inclusive variables, our findings advance ecological conservation strategies in fire-prone ecosystems.
扫码关注我们
求助内容:
应助结果提醒方式:
