Research On Building Extraction Of High Resolution Remote Sensing Images Based On Deep Learning

The rich feature information contained in high-resolution remote sensing images provides a reliable data source for the accurate extraction of buildings.At present,the research about building extraction is mainly regular buildings.Due to the large differences in the appearance and roof material of different buildings,the use of regular building extraction methods cannot accurately extract irregular buildings.This study selected Changchun,a City in Jilin Province as the study area.Besides,it selected the GF-2 satellite data as the experimental data source,and research on building extraction of high-resolution remote sensing image based on deep learning from two aspects: single-temporal remote sensing image and multi-temporal remote sensing image.The specific research content and innovation work are as follows.(1)Research on building feature extraction from high-resolution remote sensing images.Building characteristics are extremely important factors in the research of building extraction.After preprocessing the acquired high-resolution remote sensing image,the spectral characteristics of buildings were extracted by using HSI(Hue Saturability Intensity)color transformation,the texture feature from Gabor Wavelet Transform,the shape feature of combined with Graph-based and CRF,and DSBI(Difference Spectral Building Index)index feature were used to extract the building feature.The building features obtained through the above four methods have extracted spectral,shape,texture,and indices from multi-temporal high-resolution imagery.(2)Research on Building Extraction from Single-temporal high-resolution images Based on Convolutional Neural Network.GF-2 single-temporal remote sensing image was selected as the experimental data source,building samples and labels were made based on the extracted features of buildings,and then sent to CNN(Convolution Neural Networks)network and VGG(Visual Geometry Group)network for training and testing,and finally the extraction results of buildings were obtained.After the accuracy analysis of the experimental results,the average overall accuracy(OA)of regular buildings and irregular buildings of CNN network and VGG network can reach 85.15%,and the average precision can reach more than 89.9%.In addition,the average overall accuracy of VGG network is 6.7% higher than that of CNN network,so the singletemporal remote sensing image extraction results can be obtained by using VGG network.(3)Multi-temporal high-resolution image building extraction based on multifeature LSTM(Long Short Term Memory)network.Multi-temporal remote sensing imagery with time series has advantages for urban change detection and building extraction,reducing the influence of shadow and vegetation on building extraction.This study selected multi-temporal remote sensing data as data sources.Extracting the spectral,texture,shape,and index features of the multi-temporal building images,and will make sample and labels into the LSTM network to obtain the rough building extraction result,after postprocessing the accurate building extraction results will be obtained.The average overall accuracy of the final extraction result of the building can reach 93.1%,and the average precision rate can reach 90.5%.After a comparative experiment,comparing the LSTM network with the VGG network,U-Net(U-shaped Networks)network and ResNet(Residual Networks)network.The average precision of the rough extraction results of buildings on the LSTM network can reach 90.2%,which is 9% higher than the average precision of the rough extraction results of other networks.Compared with the results of building extraction of single-temporal highresolution remote sensing images,the overall accuracy of building extraction results of multi-temporal high-resolution remote sensing images was improved by 2%.This study validates the effectiveness and superiority of regular buildings and irregular buildings using multi-feature LSTM networks to extract multi-temporal images.In this paper,the single-temporal and multi-temporal high-resolution building remote sensing data was selected as experimental data.After effectively extracting the features of buildings,the multi-feature deep learning network was used to accurately extract high-resolution remote sensing image buildings,which provide reliable reference data for urban planning and design.