Landslide Recognition in High Resolution Remote Sensing Images Based on Semantic Segmentation

Abstract

In order to ensure the stable operation of high voltage transmission network, DeepLab V3+_SDF is proposed based on DeepLab V3+ for rapid and intelligent landslide detection from high resolution remote sensing images. Firstly, the backbone network is replaced by ResNet with squeeze-and-excitation (SE) attention mechanism to enhance the extraction of useful features. Secondly, astrous spatial pyramid pooling (ASPP) is reconstructed based on dense connection to expand the receptive field. More low-level features are then added to the decoder with feature pyramid networks plus (FPNP) to enhance detail recovery. Finally, a mixed loss function is proposed based on the pixel distribution to solve the sample imbalance problem. DeepLabV3+ _SDF is trained with self-made landslide remote sensing dataset. The experimental results show that the mean pixel accuracy(mPA) and mean intersection over union (mIoU) of DeepLab V3+_SDF on the landslide dataset reach 95.38 % and 85.27 %, which are 2.90 % and 7.76 % higher than those of DeepLabV3+. Finally, the trained DeepLab V3+_SDF is applied to Sichuan-Chongqing region in China, and the comparison results with manual interpretation show that the algorithm can be used for rapid identification of landslides in large-scale mountainous areas.