An optimizing dual-scale TPI framework for enhanced geomorphological classification in plateau and mountainous regions:Gaining deeper insights into topographic effects on forest fires

Accurate geomorphological classification is essential for understanding complex terrains and their influence on wildfire behavior. However, conventional methods based on arbitrarily scaled Topographic Position Index (TPI) often fail to capture multi-scale topographic heterogeneity, limiting insight into terrain–fire relationships. This study introduces an optimized dual-scale TPI framework to improve geomorphological mapping and better quantify topographic controls on burn severity. Using a 30 m DEM from southwestern China's challenging plateau-mountain region, we identified optimal TPI scale pairs (21, 23) and (53, 55) via the mean change-point method. Three machine learning models—Random Forest (RF), XGBoost, and DNN—were evaluated, with RF achieving superior accuracy (Acc = 87 %, AUC = 0.93). Feature importance analysis highlighted TPI, surface cutting depth, and slope as key factors. The quantitative analysis of burn severity across landform categories shows that mid-slope ridges exhibit the highest median RdNBR values, whereas deeply incised canyons and upland river-source areas remain within the low-severity range. This research provides a scalable method for optimal TPI selection, significantly enhancing landform classification and offering actionable insights for wildfire risk assessment and management in complex terrains. • The mean change point method identifies optimal TPI scales and addresses scale selection issues in landform classification. • Applied RF, XGBoost, and DNN models; RF achieved highest performance (Acc = 87 %, AUC = 0.93). • SHAP analysis showed that TPI dominated in mountainous and canyon landforms, while slope is the primary driver in plains. • Mid-slope ridges experienced peak fire severity, whereas canyons, headwater regions, and open slopes were low-severity zones.