A feature enhancement framework for landslide detection

Accurate landslide detection is essential for disaster mitigation and relief. In this study, we develop a feature enhancement framework that integrates attention and multiscale mechanisms with U-Net (AMU-Net) for landslide detection. The framework has four steps. First, the attention module in the convolutional block enhances the feature response of landslides when extracting high-level feature representations. Second, the multiscale module in skip connection captures more contextual information when concatenating fine and coarse features. Third, the U-Net architecture encodes feature mapping and decodes semantics representations. Fourth, the shifted window was applied to enhance the receptive field of pixels in the prediction process, which reduced the errors of landslide boundaries. Besides, we explored the effect of random split and regional split methods on model training. In the upper reach of the Jinsha River, data on unoccupied aerial vehicle (UAV) images and digital surface model (DSM) were prepared for landslide detection. The design of the framework and experiments considers the disparities in data between UAV and satellite remote sensing. The controlled experiments reported that the mean Intersection over Union (mIoU) for the proposed AMU-Net achieved 0.797, which was over 2% higher than other models. Furthermore, the visualized feature maps revealed that the proposed method can effectively restrain irrelevant feature responses in backgrounds and capture features from various receptive fields. Comparative studies on all the above experiments proved the superiority of the proposed framework for landslide detection.