Landslide Detection of 6G Satellite Images Using Multi-level Transformer Network

The 6G satellite system can help humans detect natural disasters and respond quickly through a low-altitude full-coverage network. Analyzing and identifying landslide images captured by satellites can help humans address the various hazards posed by landslides. Recently, deep learning models have been developed rapidly and demonstrated the effectiveness of landslide detection. Many models use convolutional neural networks (CNN) to extract the features of landslide images for landslide detection. However, CNN-based models cannot obtain global semantic information of images, resulting in low accuracy of landslide detection or some misjudgments. In this paper, we adopt a pre-training feature extraction network and an unsupervised multi-level transformer autoencoder for landslide detection. We first extract multi-scale features from the pre-training network, then reconstruct the image features using an autoencoder transformer network with a U-Net shape, which can better get global semantic information. We use the Bijie landslide dataset captured by satellites for experiments. The experimental results show that, compared with the original CNN model, our method can improve the detection accuracy and effectively distinguish landslide and non-landslide image data.