Automatic Landslide Mapping with Interpretable Attention-based Convolutional Neural Networks Using Remote Sensing Data

Landslide mapping plays a vital role in disaster management by providing essential information that can help decision making on mitigation and early warning strategies. However, existing automated methods often lack interpretability and miss crucial details, which limit their practical utility. This study addresses these limitations by introducing a novel Spatial Attention U-Net that leverages human visual attention to improve landslide detection and interpretability. Our proposed method integrates spatial attention modules throughout the U-Net's encoder and decoder paths, enabling the model to focus on critical image features for landslide identification. The model is trained and evaluated using a combination of high-resolution Pleiades RGB imagery, Brightness Index, and slope data. The model’s performance was evaluated using the F-1 score, precision, recall, and intersection over Union (IoU). The findings demonstrate that the Spatial Attention U-Net outperforms baseline models (U-Net, Squeeze-and-Excitation U-Net, and Channel-wise Attention), achieving F-1 scores of 73 and 79 on the testing and benchmark datasets, respectively. When applied to the inference/hold-out area, all the attention-based models outperformed the standard U-net, missing only three landslide events compared to five missed by the baseline model. Furthermore, the saliency maps reveal that the models focus on diverse regions of saliency, including edges, textures, tone, and brightness. The spatial attention U-net primarily highlights landslide edges (terrain discontinuities), while the baseline models use a mix of edges, texture, tone, and brightness. The results also indicate that dual-path attention does not lead to significant improvement in model accuracy. This approach offers a powerful tool for rapid and automated landslide mapping, indicating areas of saliency that can aid data annotation process by paying more attention to landslide object boundaries. The model interpretability further facilitates the creation of landslide inventories, especially in regions with limited ground truth data. © Geoinformatics International.