Landslide Detection Based on Differential Fusion of Multi‐Level Features From Optical Remote Sensing Images and Topographical Data

ABSTRACT Landslides are among the most devastating geohazards worldwide, causing substantial casualties and financial losses. Detecting landslides is crucial for early warning systems and disaster deformation monitoring. With the rapid advancement of Earth observation technology, optical remote sensing has emerged as one of the most widely used methods for landslide detection. However, accurately detecting new landslides using optical remote sensing images alone remains challenging due to their visual similarity to bare soil and their susceptibility to vegetation cover and human activities. As landslides typically occur in regions with specific topographical characteristics, such as slope and elevation, integrating topographical information can enhance detection accuracy. To address this challenge, this study proposes a landslide detection model (MFDF‐Net) based on differential fusion of multilevel features from optical remote sensing images and topographical data. The model incorporates a spatial feature fusion module (SP‐FFM) and a semantic feature fusion module (SE‐FFM) to effectively fuse multi‐level spatial and semantic features from these data sources. Experimental results demonstrate that MFDF‐Net outperforms mainstream segmentation models commonly used for landslide detection. The model achieves an overall accuracy (OA) of 98.73%, a Kappa coefficient of 0.7163, and a mean Intersection over Union (mIoU) of 0.7765, showcasing its superior performance in new landslide detection.