U-Net模型在大范围建筑物信息提取中的应用

APPLICATION OF U-NET MODEL IN LARGE-SCALE BUILDING INFORMATION EXTRACTION

建筑物的地理位置、面积等信息对灾害防御、应急救援及灾害评估有重要的意义。高分辨率卫星影像具有丰富的建筑物信息,可进行大范围批量化提取。使用深度学习模型对建筑物信息提取,可有效提高分割精度。为大范围获取建筑物信息,本文中以新疆伊犁州新源县为研究区,使用U-Net神经网络方法,进行批量化处理后计算出建筑物占地面积共计11160259 m^(2),占整个县域比例1.47‰,人均建筑占地面积为35.25 m^(2)。研究表明建筑物提取总体精度达到97.9%,Kappa系数达到86.4%。本研究为自动化、大范围、高精度建筑物提取工作提供了新思路。

Geographic location and area of buildings are of great significance to disaster prevention,emergency rescue and disaster assessment.High resolution satellite images have rich building information,which can be extracted in large scale and batch.Using deep learning model to extract building information can effectively improve the segmentation accuracy.Taking Xinyuan County,Yili Prefecture,Xinjiang as the research area,this paper uses U-Net neural network method to extract building information in a wide range.Through batch processing of the data in the study area,it is calculated that the total area covered by buildings is 11160259 m^(2),accounting for 1.47‰of the whole county,and the per capita building area is 35.25 m^(2).The research shows that the overall accuracy of building extraction is 97.9%,and the kappa coefficient researches 86.4%.This study provides a new idea for automatic,large-scale and high-precision building extraction.