Mapping taluses using deep learning and high-resolution satellite images

Taluses are widely distributed in alpine regions such as the Tibetan Plateau. Despite their critical environmental and geohazard roles, taluses have only been mapped in limited regions. This study presents an effective approach to identifying taluses by capturing their morphological features using deep learning (DeepLab V3+ with an attention mechanism) and high-resolution satellite images. The approach was applied to 2-m-resolution GaoFen satellite images to map taluses in the source area of the Yellow River in the eastern Tibetan Plateau and compile the first comprehensive talus records of the region. The results obtained were highly accurate, with 88.6% of the F1 score compared to manual interpretations. The mapped taluses covered approximately 3.89 × 103 km2, 3.19% of the region, with the vast majority located at moderate elevations (4,000–5,000 m asl) and moderate slopes (10–35°). The mapped areas are characterized by frequent freeze – thaw cycles, significant terrain ruggedness, and sparse vegetation cover. The results reveal other interesting characteristics of talus distribution regarding lithology, permafrost thermal stability, and precipitation. These findings could provide valuable information about the forces that drive talus formation and improved understanding of taluses in the Tibetan Plateau and globally by combining recent advances in artificial intelligence and Earth observations.