[M-AG33] Satellite Land Physical Processes Monitoring at Medium/High/Very High Resolution

[MAG33-P06] A universal adapter in segmentation models for transferable landslide mapping using Sentinel-2 satellite images

Ruilong Wei^1,2, *Yamei Li³, Yao Li⁴, Jiao Wang¹, Chunhao Wu¹, Shunyu Yao^5,6, Chengming Ye⁷ (1. Key Laboratory of Mountain Hazards and Earth Surface Process, Institute of Mountain Hazards and Environment, Chinese Academy of Sciences, 2. University of Chinese Academy of Sciences, 3.State Key Laboratory of Tibetan Plateau Earth System, Environment and Resources (TPESER), Institute of Tibetan Plateau Research, Chinese Academy of Sciences, 4.State Key Laboratory of Hydroscience and Engineering, Tsinghua University, 5.China Institute of Water Resources and Hydropower Research, 6.Research Center on Flood and Drought Disaster Reduction of the Ministry of Water Resources, 7.Key Laboratory of Earth Exploration and Information Technology of Ministry of Education, Chengdu University of Technology)

Keywords:Landslide mapping, Deep learning, Tibetan Plateau, Sentinel-2 satellite images

Landslides are widespread and devastating disasters in mountainous areas. Efficient landslide mapping is crucial for disaster mitigation and relief. Recently, deep learning methods have shown promising results in landslide mapping using satellite imagery. However, the sample sparsity and geographic diversity of landslides have challenged the transferability of deep learning models. This study proposed a universal adapter module that can be seamlessly embedded into existing segmentation models for transferable landslide mapping. The adapter can achieve high-accuracy cross-regional landslide segmentation with a small sample set, requiring minimal parameter adjustments. In detail, the pre-trained baseline model freezes its parameters to keep learned knowledge of the source domain, while the lightweight adapter fine-tunes only a few parameters to learn new landslide features of the target domain. Structurally, we introduced an attention mechanism to enhance the feature extraction of the adapter. To validate the proposed adapter module, 4321 landslide samples were prepared, and the Segment Anything Model (SAM) and other baseline models, along with four transfer strategies were selected for controlled experiments. In addition, Sentinel-2 satellite images in the Himalayas and Hengduan Mountains, located on the southern and southeastern edges of the Tibetan Plateau was collected for evaluation. The controlled experiments reported that SAM, when combined with our adapter module, achieved a peak mean Intersection over Union (mIoU) of 82.3 %. For other baseline models, integrating the adapter improved mIoU by 2.6 % to 12.9 % compared with traditional strategies on cross-regional landslide mapping. In particular, baseline models with Transformers are more suitable for fine- tuning parameters. Furthermore, the visualized feature maps revealed that fine-tuning shallow encoders can achieve better effects in model transfer. Besides, the proposed adapter can effectively extract landslide features and focus on specific spatial and channel domains with significant features. We also quantified the spectral, scale, and shape features of landslides and analyzed their impacts on segmentation results. Our analysis indicated that weak spectral differences, as well as extreme scale and edge shapes are detrimental to the accuracy of landslide segmentation. Overall, this adapter module provides a new perspective for large-scale transferable landslide mapping.