Enhancing deep learning-based landslide detection from open satellite imagery via multisource data fusion of spectral, textural, and topographical features: a case study of old landslide detection in the Three Gorges Reservoir Area (TGRA)

Abstract

Deep learning (DL) techniques are now predominant in landslide detection, aiding in risk assessment and disaster management. Detecting old landslides from open satellite imagery is challenging due to their altered shapes over time and minimal variation from the natural environment. Additional, DL models often function as unexplained black boxes. This study employs a DL classifier based on the ENVINet5 architecture, combined with a multisource integrated dataset, to detect old landslides from Sentinel-2 images. Gradient-weighted class activation mapping (Grad-CAM) is used for visual explanation of the classifier. Experimental results from the Three Gorges Reservoir Area (TGRA) indicate that the ENVINet5-based model, utilizing spectral, textural, and topographical features, achieves outstanding performance with an accuracy of 0.9589 and an F1-score of 0.9105. Multisource data fusion significantly enhances model performance, improving accuracy by up to 22.6% compared to using only spectral data. Grad-CAM visualizations confirm the model accurately identifies landslide boundaries of various shapes. The proposed method offers improved performance, visual explanations, cost efficiency, and accessibility, making it suitable for detecting old landslides from open satellite imagery.

Keywords

• Landslide detection
• multisource data fusion
• ENVINet5
• open-source satellite data
• gradient-weighted class activation mapping (Grad-CAM)