Deep Learning Framework Integrating Self-Supervised Learning and Attention for Landslide Susceptibility Mapping

The field of landslide susceptibility mapping has long faced a lack of high-precision labeled data. This paper proposes a deep learning framework that combines self-supervised learning with an attention mechanism. The framework utilizes the classic DeepLabv3+ model as its backbone and pre-trains it using a self-supervised learning strategy to enhance feature representation. At the same time, it introduces the CBAM module to enhance the model's sensitivity to spatial and channel information in the input features, thereby strengthening its ability to extract information from remote sensing data. This paper conducts standardized comparative experiments between the proposed deep learning framework and several classical models commonly used in landslide susceptibility studies. The results show that the proposed method achieves an F1 score of 0.820 and a mIoU of 0.846 on the test set, and it outperforms the comparison models overall. In addition, ablation studies are carried out to demonstrate that both self-supervised pre-training and the introduction of the attention mechanism contribute positively to improving model accuracy. Finally, the paper extracts and analyzes the channel attention scores from CBAM, which reveal the relative importance of different influencing factors in the proposed model. The findings of this study provide a new framework for applying deep learning models to regional landslide susceptibility analysis. Furthermore, the proposed framework shows the potential to improve performance by making effective use of limited high-precision labeled data.