Intelligent Recognition Of Co-seismic Landslides Based On Deep Learning

Rapid mapping of co-seismic landslides inventories is essential for emergency management and loss assessment.Currently,deep learning algorithms follow a traditional supervised learning workflow and then the trained model is used to predict landslides in surrounding areas,achieving end-to-end,pixel-level landslide semantic segmentation with high accuracies.When there is a new study area or unknown scenarios landslide extraction task,due to the different characteristics of earthquake landslides,vegetation features and landform features in different scenarios,the performance of model trained only for a specific dataset will be greatly reduced in the new environment.This paper conducts experiments around the problem of generalizability of co-seismic landslides across scenarios,taking regions of Iburi,Jiuzhaigou,Palu and Haiti as study area,and the freely available Sentinel-2 imagery and digital elevation model(DEM)were used.The main research content of the paper are as follows:(1)Considering that different geomorphic units make multi-channels have different contributions to the occurrence of landslides,we introduced the attention mechanism into FCN,which made the data adaptively assigns different importance weights to the impact factors to improve the robustness of the model.We first proposed a new convolutional neural network model,called Res U-SENet to generate a segmentation map of landslides and its performance was compared with traditional convolutional neural network model.(2)Multi-domain detectors were constructed by mixing samples from Iburi and Jiuzhaigou in different ratios and the performance of multi-domain detectors were explored.It was concluded that the mixing ratio of the samples from the foreign domain and domain of interest directly determined the detection performance of the model.When foreign domain samples are a minority,it helps to improve the performance of the detector in the domain of interest.When combining half number of foreign domain samples(50%)with half of the local samples(50%)also maintains almost the same detection performance as if trained entirely with local samples.The proposed multi-domain detector effectively reduces the number of new samples for model training in landslide disaster emergency mapping.(3)We further used detectors trained by Jiuzhaigou and Iburi samples directly on completely unknown domains,the F1-Score under the Res U-SENet model reached 0.6875 and0.6916 in Palu and Haiti respectively,proving our detectors have a good generalization performance.