Application of multi-scale deep convolutional neural networks (MDCNN) to susceptibility assessment of rainfall-induced landslides in typical southern hilly areas: A case study of Wanxiu District, Wuzhou City, Guangxi, China

Rainfall-induced landslides are governed by a combination of factors, including topography, geology, rainfall, and human activity. Traditional landslide susceptibility assessment models often struggle to capture the complex spatial non-linear relationships among environmental factors, thereby limiting the accuracy of regional landslide susceptibility predictions. Taking Wanxiu District, Wuzhou City, Guangxi, as the study area, this research utilized eleven environmental factors—surface incision depth, slope gradient, aspect, Digital Elevation Model (DEM), topographic relief, plan curvature, profile curvature, lithology, Topographic Wetness Index (TWI), land use, and rainfall—to construct a Multi-scale Deep Convolutional Neural Network (MDCNN) for assessing rainfall-induced landslide susceptibility. The model's performance was compared with Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and SegFormer. The results indicate that slope gradient, surface incision depth, topographic relief, lithology, and rainfall are the key controlling factors for landslides in the study area. Specifically, surface incision depth exhibited the highest frequency ratio (FR = 1.149), rainfall showed the highest contribution rate in the Random Forest model (0.1422), and slope gradient demonstrated the greatest spatial explanatory power (q = 0.1334). Landslide susceptibility in the study area displayed distinct spatial heterogeneity; zones of high and extremely high susceptibility were primarily distributed along the valley slopes and hilly terrain near the confluence of the Gui River and the Xi River. The extremely high susceptibility zone covered an area of 21.92 km² (3.42% of the total study area) and encompassed 318 historical landslides, representing 71.6% of the total recorded landslides. Comparative analysis revealed that the MDCNN model achieved the highest prediction accuracy (AUC = 0.966)—outperforming LR (0.768), SVM (0.790), RF (0.908), and SegFormer (0.890)—demonstrating its ability to effectively learn complex spatial non-linear relationships among environmental factors and accurately identify areas with high landslide susceptibility. Research indicates that the MDCNN model offers a highly accurate, stable, and regionally applicable technical method for assessing the susceptibility to rainfall-induced landslides in the hilly regions of southern China, providing a scientific basis for landslide risk identification, monitoring and early warning, and disaster prevention and mitigation in territorial spatial planning. © 2026 The Authors.