Classification And Statistics Of Buildings Based On UAV Images And Deep Learning Technology

Pre-assessment of earthquake disaster losses in areas where earthquakes may occur is an important measure to improve the level of earthquake emergency preparedness in risk areas and the accuracy of rapid post-earthquake assessment.In the process of pre-assessment,it is necessary to investigate the local geographical features,housing construction,traffic roads,secondary disasters and other matters.How to accurately collect data and determine the structure type of a large area of buildings in a short time is an important link in the pre-evaluation work.For many years,the Institute of Geology of China Seismological Bureau has put forward and developed the lethality assessment method in order to meet the needs of pre-assessment,and one of the important parts of this method is to evaluate the lethality level in various regions,and in the composition of this level,the structure,seismic capacity and number proportion of buildings occupy a very important position.UAV has a natural advantage in meeting the above needs.This paper will explore the technical methods to obtain building data quickly and accurately with UAV images based on the actual survey of multiple ground.In this paper,Xuyi County,Huai'an City,Jiangsu Province is selected as the study area,and the remote sensing images taken by small UAV are used as the data source,trying to apply deep learning technology to automatically interpret the distribution of building structures in the area.In the experiment,the UAV data of forty towns and streets in Xuyi County were collected to determine the classification basis of structural types of houses in the area;Then Xuyi County under the jurisdiction of the four market towns and four weak villages of a total of eight research points as a test area,through the UAV digital images,using object-oriented image analysis method to extract the image of housing construction;Finally choose deep learning framework Pytorch for classification model modeling,using the Res Net network model,the extracted buildings are classified according to the house structure,the finalstudy area overall house structure distribution.The field research data is used as the real data to evaluate the classification accuracy of the results obtained by the UAV remote sensing method.The overall classification accuracy of the eight sites is above75%.Through the study of this paper,it shows that the UAV low-altitude remote sensing technology in the identification of the housing structure of the county has a certain feasibility,and it can provide some help for earthquake pre-assessment investigation.The classification accuracy can be further improved by improving the house extraction algorithm and expanding the house training set.This shows that UAV technology has good research prospects in pre-evaluation work.