Multimodal Deep Learning for Landslide Detection Using Multispectral Satellite Imagery and Topographic Features

Accurate landslide detection is important for reducing disaster risk, supporting emergency response, and improving land management. In recent years, deep learning has become an effective tool for extracting landslide information from remote sensing imagery. However, combining spectral and topographic data remains a challenge, particularly when working in complex terrain. This study investigates the use of a deep learning model for landslide detection using the Landslide4Sense dataset, which contains Sentinel-2 multispectral imagery together with topographic information. A U-Net model was trained using six input channels, including RGB bands, the Normalized Difference Vegetation Index (NDVI), Digital Elevation Model (DEM), and slope. These data provide complementary information about land cover, vegetation, and terrain characteristics that are closely related to landslide occurrence. Before training, the data were normalized; NDVI was calculated from the multispectral bands; and missing values were handled. The model was trained using the Adam optimizer with binary cross-entropy loss and evaluated with common segmentation metrics. The proposed model achieved a validation accuracy of 89.9%, with an F1-score of 0.71, precision of 0.79, and recall of 0.64. These results provide a strong baseline for multimodal landslide segmentation. As part of the ongoing research, the performance of U-Net will be compared with more recent deep learning models, including transformer-based architectures and other semantic segmentation networks, to assess their accuracy, robustness, and computational cost. This comparison will help identify the most suitable models for automated landslide mapping from multispectral satellite imagery. The findings are expected to support the development of reliable deep learning methods for landslide detection and provide useful guidance for future remote sensing applications in natural hazard monitoring.