Landslide Image Segmentation with Attention Residual U-Net: A Hybrid Deep Learning Model

Landslides are among most destructive geohazards, rising due to human activity and climate change, posing significant risks to lives and economies. Rapid assessment after a landslide is crucial for directing emergency resources, requiring precise detection and segmentation of affected areas. Remote sensing, combined with deep learning, has become an effective tool for monitoring these hazards. However, the diverse characteristics of satellite images often lead to false detections and increased computational costs, limiting model accuracy. This limitation arises from the inability of current models to effectively capture detailed features in complex, multi-scale backgrounds and datasets that vary significantly across regions. To address these challenges, the authors present an Attention Residual U-Net (ARU-Net), an advanced deep learning framework that integrates residual connections and attention mechanisms within the basic U-Net structure to enhance feature extraction and segmentation accuracy. Attention mechanisms, a key technique in computer vision, enhance feature extraction and improve segmentation performance by focusing on relevant parts of the input data, while residual connections improve learning efficiency providing detailed feature refinement despite their higher computational costs. Additionally, the binary focal loss is employed to increase the model’s capacity to handle intricate landslide characteristics, particularly in unbalanced data. Experimental results indicate that the suggested ARU-Net significantly improved overall segmentation accuracy compared to Attention U-Net and basic U-Net, utilizing satellite imagery from Luding County as input. The segmentation results are verified through the Jaccard similarity index (JSI), Accuracy (ACC), and mean Intersection over Union (mIoU). The suggested ARU-Net is observed to achieve an accuracy of 95.16% and mIoU of 83.31%, respectively. Additionally, our model outperforms the U-Net baseline presented in the study by Wang et al., improving accuracy by 5.13% and mIoU by 11.63%. These findings indicate that the suggested ARU-Net provides a robust and effective solution for automatic landslide segmentation, contributing to better disaster response and risk mitigation.