Landslide Detection of High-Resolution Satellite Images using Asymmetric Dual-Channel Network
Yaohui Liu, Wenzhuo Zhang, Xiaoxian Chen, Mingyang Yu, Yingjun Sun, Fei Meng, Xiwei Fan
Abstract
Landslide detection from high-resolution satellite imagery plays a significant role in disaster management. Recently, deep learning has emerged as one of the most powerful tools for landslide detection. However, the existing deep learning models for landslide detection still have room for improvement. In this paper, we propose a novel deep learning model named DCA-Net for the automatic detection of landslides. This model adopts an asymmetric encoder-decoder structure. The core module designed in the DCA-Net is the dual-channel depthwise block, integrating two parallel channels of asymmetric depthwise separable convolutions and residual connections to enlarge the receptive field and improve the performance. Experiments are conducted on the open-source Bijie landslide dataset. Several state-of-the-art networks are also employed for quantitative and qualitative comparisons. The results indicate that our proposed DCA-Net is superior to other deep learning models, which can be well utilized in landslide detection from aerial images.