Building Extraction Method Based On Deep Learning For Post-Earthquake In Remote Sensing Images

Due to the suddenness and destructiveness of earthquake disasters,earthquakes will cause serious economic losses and casualties in the affected areas.Buildings are the main places for human activities,so timely and accurate detection of the disaster-affected areas of buildings after an earthquake is of great significance for the formulation of emergency plans,post-disaster relief,and post-disaster reconstruction.Relying on traditional ground surveys and historical data to obtain information on disaster-affected areas of buildings has high accuracy and credibility,but there are disadvantages such as untimely information statistics,heavy workload,low efficiency,and high cost.With the continuous development of remote sensing technology and the substantial improvement in the performance of imaging equipment,the resolution of remote sensing images is getting higher and higher,the coverage is wider and wider,and the imaging cycle is getting shorter and shorter.High-resolution optical remote sensing images after earthquakes are fast and accurate.The ground detection of building disaster areas provides strong data support.In sudden earthquake monitoring,the application of remote sensing technology has obvious advantages.At the same time,thanks to the rapid development of deep learning technology in recent years,technologies such as deep convolutional neural networks can automatically learn advanced feature expressions in remote sensing image data,so that the application of post-earthquake high-resolution optical remote sensing images for building Automatic detection of disaster areas is possible.Therefore,this paper combines remote sensing images with cutting-edge computer deep learning technology,applies convolutional neural network technology to earthquake disaster scenes,and extracts the disaster information of buildings after the earthquake,in order to improve the accuracy and automation of the detection of the disaster area of buildings after the earthquake.The main research work of this paper is as follows:(1)Construct a data set of damage to buildings after the earthquake.Aiming at the lack of a complete data set of damaged buildings after the earthquake during model training,analyze and study the characteristics of building damage information on remote sensing images and the classification standard of damage degree,through label annotation,image cropping,image enhancement and random image division.Construct a data set containing three types of labels: undamaged buildings,damaged buildings,and background,so as to prepare data for later model training and performance evaluation.(2)Constructing the BR-Mask R-CNN for the detection of damaged areas of buildings after the earthquake.Aiming at the problems of incomplete feature extraction and rough boundary of the extraction results commonly found in existing methods,based on the Mask R-CNN with better instance segmentation effect,this thesis researches and improves its backbone feature extraction network and Mask output branch,thus constructing a new type of post-earthquake building damage area detection model,improving the accuracy of post-earthquake damage area detection.Although this model improves the accuracy of detection results,due to its high computational complexity and the need for more storage and computing resources,this thesis continues to study a more efficient detection model for post-earthquake building disaster area.(3)Construct the ELRNet for the detection of damaged areas of buildings after the earthquake.In order to improve the timeliness of building disaster-affected area detection,this thesis is inspired by the fact that the multi-layer feature extraction module in the improved Mask R-CNN can effectively improve the feature extraction ability and some excellent lightweight network design strategies,continues to explore and study a more efficient and lightweight post-earthquake building disaster area detection model ELRNet.The core of ELRNet is the multi-level self-designed lightweight feature extraction module LFEM,which improves the accuracy and efficiency of the detection of disaster-affected areas of buildings after the earthquake.