Research On Disaster Information Extraction Of Roads And Bridges From Remote Sensing Images

After a major earthquake,it is very important to obtain the macroscopic scope of the disaster-affected areas and the damage of the main disaster-bearing body in time.Among them,the rapid acquisition of the damage location,scope and degree of damage to typical elements of lifeline projects such as roads,bridges,dams,etc.is an important guarantee for the efficient execution of the entire rescue work,and it is also one of the main basis for decision-making support for post-disaster reconstruction planning.Compared with field surveys,the remote sensing can quickly and comprehensively obtain disaster information about lifeline engineering elements.The effective extraction of disaster information of the post-earthquake lifeline project from remote sensing images is of great significance for supporting earthquake relief and effectively improving rescue efficiency.However,the traditional remote sensing estimation method only uses some artificially designed shallow features in the image,which has great uncertainty and is complicated and time-consuming.For the post-earthquake remote sensing image of high spatial resolution with complex background,the accuracy of model is often insufficient as well as speed.Thus it is difficult to achieve the desired effect.In view of this,with the support of National Key Research and Development Program "Research on Key Technologies of Rapid Disaster Investigation after Great Earthquake Coordinated by Satellite,Aircraft and Ground"(Project NO.2017YFC1500902),this paper carried out research on the acquisition of disaster information of typical lifeline projects after disasters based on knowledge rules and deep learning.The disaster information features of roads,bridges were extracted and analyzed by using high-resolution remote sensing image data after the earthquake.This research established corresponding knowledge rules and proposed a knowledge-based model for acquiring disaster information of typical lifeline engineering elements.Furthermore,the research explored the feasibility of deep learning in acquiring disaster information of lifeline engineering.The main research achievements and innovations of this paper include the following three aspects:(1)The knowledge-based IFCM-SVM remote sensing information extraction model of post-earthquake road disaster was studied and put forward.Different ground objects have different features in remote sensing images.In the process of target detection and recognition,the research and analysis of their features in remote sensing images is a very necessary and key step.Through the research and analysis of such feature differences,damage information can be better extracted by combining with the model.This paper analyzes in detail the image segmentation theory,fuzzy classification theory,intuitionistic fuzzy set and support vector machine method and principle involved in the road disaster extraction model.In this paper,a method of disaster situation extraction based on first segmentation and then classification is proposed,which can effectively transform multiple classification problems into binary classification problems.Aiming at the remote sensing image after the disaster,an improved FCM algorithm based on intuitionistic fuzzy set was proposed to segment the image,and the interested layer containing the disaster target was obtained.The principle of IFCM algorithm was described in detail,and the advantages of IFCM compared with FCM algorithm were verified by experiments.For the extraction of road disaster information,it is assumed that the extraction of road disaster information is completed through the IFCM-SVM model without the support of road vector data,and it is verified by the accuracy evaluation results of experimental data.The experimental results show that the knowledge-based IFCM-SVM model can better extract road disaster information.(2)A remote sensing information extraction model for bridge disaster situation after earthquake based on spatial relationship is proposed.By analyzing the characteristics of the bridge in the post-earthquake image,the contextual characteristics of the bridge and the surrounding environment are comprehensively utilized.In particular,according to the spatial relationship between the bridge and the river in the remote sensing image,a bridge disaster detection model based on the spatial relationship is proposed,and the implementation process of the algorithm is described in detail.The spectral value of water body was enhanced by building index so as to expand the background difference between water body and nonwater body.The iterative threshold method was used to extract rivers,and the layer containing the bridge target was obtained by using IFCM algorithm.The candidate Bridges were extracted by combining the two methods and the disaster analysis was carried out on them.According to the experimental results,the bridge disaster extraction model based on spatial relationship can effectively extract the disaster information of bridge.(3)The target sample set of post-earthquake road disaster is constructed,and the technical method and process of post-earthquake road disaster information extraction based on deep learning are proposed.A sample set of road disaster target detection was constructed based on the postearthquake high-resolution remote sensing images with the slice size of 416×416.Through enhancement,a total of 2817 sample images were collected,and each sample image was labeled with two types of road disaster situations: slight damage and severe damage,and a total of 1,955 road disaster targets were labeled.For visual observation,the disaster information of roads and Bridges in the remote sensing images after the earthquake is usually small-scale information.Compared with other deep learning models,the fast and real-time YOLOV3 model has a higher detection accuracy for small targets,and YOLOV3-S-GIOU model has an excellent performance in the detection of post-earthquake building damage information.In view of this,this paper attempts to use the YOLOV3-S-GIOU detection model proposed by our research group for collapsed buildings after the earthquake to carry out target detection on the disaster situation of roads after the earthquake.The core of the model is to use Shufflenet V2 lightweight network to replace Darknet53 and effectively reduce network parameters.Secondly,the XY loss at the center of the prediction box and the WH loss at the width and height in the loss function are uniformly replaced with GIOU loss.In this study,the constructed road disaster target samples and bridge samples were input into the model for training,and the experiment was carried out with "8.8" Jiuzhaigou earthquake images of 0.5m.The extraction accuracy of the model for road severe damage is 82.2%.