PACDNet: a landslide extraction network based on parallel attention

Landslides pose a serious threat to the economy and human safety. Rapid and efficient landslide extraction is essential for post-disaster relief and reconstruction efforts. Current landslide extraction techniques still face challenges in accurately identifying landslides boundaries. To address this issue, we propose the Parallel Attention Change Detection Network (PACDNet), a network designed to extract landslides by calculating changes between pre-disaster and post-disaster images. Our method is an encoder-decoder network that utilizes a Siamese network structure with StarBlock from StarNet in the encoding stage, allowing for higher feature extraction capability with lower computational costs. In the decoding phase, the Parallel Attention Module (PAM) enhances the model's landslide boundaries extraction capability by simultaneously computing spatial attention and channel attention. To validate the effectiveness of our proposed PACDNet, we used the Global Very-High-Resolution Landslide Mapping (GVLM) dataset. This dataset includes pre-disaster and post-disaster images of 17 different landslide events. After comparing our model with four recently proposed models, PACDNet demonstrated improvements of 2.31% and 3.41% in Intersection over Union (IoU) on the Kaikoura and Tbilisi datasets, respectively. This indicates that PACDNet has a significant advantage in landslide extraction across different landslide events.