Hybrid attention residual U-Net++ with Atrous Spatial Pyramid Pooling (HARU-ASPP) for landslide segmentation

Landslide segmentation is a key tool for identification, monitoring, early warning, and disaster management. However, landslide segmentation using remote sensing imagery faces challenges due to complex textures or often blurred landslide boundaries. Machine learning and deep learning methods have been able to improve accuracy, but still struggle to balance fine spatial details with extensive contextual information. To overcome these limitations, this study proposes an architecture for landslide segmentation, namely Hybrid Attention Residual U-Net++ with Atrous Spatial Pyramid Pooling (HARU-ASPP). The proposed model integrates a residual convolutional block mechanism, an ASPP module, and a hybrid attention module (spatial and channel), with nested skip connections in the U-Net++ architecture. This model is performed to enhance the feature preservation, context understanding, and boundary accuracy. In this regard, the evaluation of HARU-ASPP incorporates Landslide4Sense as the benchmark dataset by using metrics, including mean Intersection over Union (mIoU), Jaccard index, mean Average Precision (mAP), Accuracy, F1-score, Precision, and Recall. As a result, HARU-ASPP outperforms other methods, achieving better Precision–Recall balance, as well as mIoU (80.45 %) and mAP (84.61 %). By combining hybrid attention, residual learning, and multiscale context aggregation, the experimental results demonstrate an effective technique for producing reliable landslide segmentation on complex remote sensing imagery.