Improved DeepLabV3+model for landslide identification in high-resolution remote sensing images after earthquakes
Postearthquake emergency rescue must use the method of"deep learning+remote sensing"to identify landslides from high-resolution remote sensing images quickly after an earthquake.However,an excellent deep learning model cannot be separated from a large-scale,high-quality dataset as a support.The size of the dataset and the quality of the target category data annotation information contained in it directly affect the performance and application effect of the deep learning model.So far,a few public datasets of deep learning landslide identification are available;however,they hardly meet the task requirements of researchers who use deep learning methods to conduct landslide identification.Thus,this study uses the visual interpretation method to annotate the pixel level of GF-6 remote sensing image and Google Earth image taken after the earthquake and uses the digital elevation model and optical image data to establish a three-dimensional model to assist and ensure the accuracy of landslide annotation.Finally,a public deep learning landslide dataset with a spatial resolution of 2 m is established;it can be used to train deep learning semantic segmentation and target detection models.The dataset contains 11581 groups of data,among which the training set contains 9265 groups of data;the validation set contains 1158 groups of data,and the test set contains 1158 groups of data,which far exceeds the data volume of the existing public landslide dataset and basically meets the training data volume requirements of most deep learning landslide identification tasks.In addition,to recognize the boundary and detail information of landslides better and improve the accuracy of landslide recognition,this study proposes an improved DeepLabV3+landslide recognition model based on DeepLabV3+network by introducing channel attention mechanism,feature map fusion module,and transposed convolution.The feature map fusion module of the channel attention mechanism is used to adjust the weight between different feature map channels in the model training process to fuse effectively the low-order and high-order feature output by the coding structure.The transposed convolution uses a learnable convolution kernel to upsample the feature map to process complex image structure and semantic information well.During model training,this study uses the transfer learning method to transfer the backbone network architecture and its parameters of the ResNet50 model trained on the ImageNet dataset to the encoder structure of the DeepLabV3+network model to accelerate the training of the model.Experimental results show that compared with the mainstream algorithms(FCN,U-Net,SegNet,DeepLabV3+),the improved DeepLabV3+model has better extraction effect on the boundary and details of the landslide,and the results are closer to the real label;among them,MIOU is 87.24％,recall is 92.47％,precision is 90.35％,Fl score is 90.87％,and pixel accuracy is 98.91％.The code and data for this article are available at https://github.com/ZhaoTong0203/landslides_identification_model_code.git.This research provides robust support for the advancement of deep learning in landslide identification and offers substantial practical assistance for postearthquake emergency rescue efforts.
high-resolution remote sensing imageslandslide data setsdeep learningDeepLabV3+GF-6
万方数据
维普
