Research On Extraction Method Of Road Damage Information From High-resolution Remote Sensing Image Based On Deeplabv3+

The earthquake and its secondary disasters have caused great harm to human society.The destruction of road affects the on-site disaster investigation,the release of disaster relief materials and rescue work after the earthquake.Therefore,it is crucial to quickly and efficiently obtain road damage information in the disaster area after the earthquake.On the one hand,the remote sensing images of the disaster area can be quickly and efficiently acquired because its technology is not restricted by ground conditions;on the other hand,with the development of remote sensing technology,we can obtain more and more high-resolution satellite and aerial image data,which can be more cleariy reflects the road damage information.At present,many domestic and foreign researches on extracting road damage information based on post-earthquake remote sensing images have been carried out.Although traditional visual interpretation methods are highly accurate,they cannot meet the requirements of post-earthquake emergency response timeliness because they take a lot of time and manpower.The computer-aided road seismic damage extraction method can efficiently obtain the macroscopic and comprehensive information after the earthquake,but its level of automation,accuracy and migration are still far from the actual application requirements.Deep learning can automatically establish the extraction model of the research object according to the characteristics of the image,and is successfully applied in the field of image processing.However,there are few researches on the application of deep learning to road disaster extraction after the earthquake,and the existing deep learning methods cannot make full use of the image context information and the imbalance of sample categories.Therefore,the extraction of remotely sensed road damage information based on deep learning has important theoretical and practical value for the acquisition of road damage information after the earthquake.Based on the sample data sets of the deep learning model,the paper builds the Deeplabv3+ road extraction model,and studies and analyzes the impact of training times and training sample size on the performance of the model;then experiments results show Deeplabv3+ has a higher accuracy by comparing with the currently useddeep learning full convolutional neural network.Then the trained Deeplabv3+ model is applied to the extraction of road damage information from the GF-1 satellite image with 2 meters resolution acquired after the Ludian earthquake and the UAV image with 0.2 meters resolution acquired after the Lushan earthquake.Experiments have proved that the Deeplabv3+ road damage extraction model based on high-resolution remote sensing images has good accuracy and mobility,a which can provide effective support for earthquake relief.The main research contents of this paper are as follows:(1)Summary and analysis of methods for extracting damaged road information from high-resolution remote sensing images after the earthquake.According to domestic and foreign literature investigations,it is found that the automatic or semi-automatic road seismic damage information extraction algorithms and technologies that have been proposed have achieved certain results and overcome the problem of timeliness to a certain extent.However,these methods are used in remote sensing image road extraction,in the process of image segmentation,the image patches obtained by image segmentation often cannot fully consider the context information of the image,the road damage information cannot be identified more accurately,and the accuracy needs to be further improved.On this basis,the application research of deep learning methods in the extraction of intact roads is analyzed.In view of the current problems,the high-scoring road damage information extraction based on the Deeplabv3+ model is proposed method.(2)Road information extraction experiment based on Deeplabv3+ model and FCN model.First,by comparing and analyzing the commonly used deep learning networks theoretically,the Deeplabv3+ model performs better pixel-level semantic segmentation of images;then model training sample data sets are built from 236 scenes with a full-color resolution of 3.2 meters and a multi-spectral resolution of 0.8meters BJ-2 Satellite images,which are composed of 10754 pairs of 3-channel image pairs with a size of 500500? pixels for model training and testing,using the samples sets,the Deeplabv3+ model and FCN model are built,and the impact of the training sample size on the model performance are further analyze.The total accuracy of the Deeplabv3+ model is 96%,and its Kappa coefficient is 0.866.The total accuracy of the FCN model is 83.6%,and its Kappa coefficient is 0.734.The experiment provesthat the Deeplabv3+ model has higher accuracy.(3)The road damage information extraction experiment based on the Deeplabv3+ model from multi-source remote sensing image with high resolution.Taking the Lu Dian earthquake and Lu Shan earthquake as examples,respectively,the trained Deeplabv3+ model's extraction accuracy of damaged roads and the mobility of the model were tested.23 damaged roads were visually interpreted and 25 damaged roads were extracted using the Deeplabv3+ model,in the GF-1 satellite image of Longjing Village in Ludian County with 2 meters resolution after the Lushan earthquake.The accuracy of Deeplabv3+ model extraction was 83%.Based on the UAV images with 0.2-meter resolutionin Baosheng town after the Lushan earthquake,the road was 5057.5 meters by visual interpretation,the undamaged road is 4018 meters,the damaged road is 446.7 meters extracted using the Deeplabv3+ model.The model road extraction accuracy is 88%.The Deeplabv3+ model has better accuracy in the extraction of damaged roads from high-resolution remote sensing images with lower resolution and higher resolution than the training samples.Therefore the model has better migration.