Cross-Branch CNN-MLP Integration for Improving Landslide Spatial Probability on Mt. Umyeon, Korea

Abstract

Accurate maps of landslide spatial probability (LSP) are crucial for planning and risk reduction in steep, urbanized areas. A hybrid Convolutional Neural Network (CNN) - Multilayer Perceptron (MLP) model is introduced for mapping LSP in Mt. Umyeon, Korea. The design combines two complementary views of each location: a convolutional branch learns spatial context from multi-channel image patches. At the same time, a multilayer perceptron captures local numeric and categorical attributes at the central point. Feature-level fusion concatenates the embeddings from both branches and feeds a lightweight classifier to produce probabilities. Performance was assessed under consistent data splits and training protocols. Training AUCs reached 0.887 (MLP), 0.894 (CNN), and 0.903 (CNN-MLP). More importantly, validation AUCs were 0.808 (MLP), 0.821 (CNN), and 0.854 (CNN-MLP), indicating stronger generalization for the fused representation. These gains reflect the complementary nature of neighborhood structure learned by the CNN and pointwise information stabilized by the MLP. The results show that a compact, feature-level fusion of CNN and MLP can materially improve spatial probability mapping of landslides. The approach provides a practical route to more reliable probability surfaces for decision support in mountainous urban regions.