Research On Building Extraction From High-resolution Remote Sensing Images Based On Convolutional Neural Network

As an important artificial feature,buildings are one of the important indicators of urbanization and economic development.As a kind of spatial big data,high-resolution remote sensing image can reflect the rich details of buildings and provide the possibility of building extraction.However,the improvement of image resolution does not mean the improvement of image interpretation capability and building extraction accuracy.The artificially designed feature extraction method has been unable to meet the needs of extracting buildings from massive remote sensing images.Therefore,studying the idea of using convolutional neural networks to automatically extract features to improve the accuracy of building extraction has important application value and practical significance.Aiming at the problem of building extraction with high-resolution remote sensing image,the main research contents and achievements of this paper are as follows:Firstly,in view of the small number of high-resolution remote sensing image building data sets published on the Internet,a high-resolution remote sensing image building data set was established.Using the high-resolution No.2 image with a resolution of 1 meter as the data source,the buildings in the main city of Chongqing are selected by hand-drawing to make a building dataset covering an area of about 60 square kilometers,and the data is passed The enhanced method enhances and expands the data set for use by the convolutional neural network in training..Secondly,an automatic extraction network of densely connected U-Net buildings is proposed.On the basis of U-Net,the dense connection is introduced into the network’s downsampling and upsampling.The dense connection increases the flow of information between layers,strengthens the capability of feature reuse,and improves the accuracy of building extraction.In the experimental comparison of the three building data sets of Mnih,ISPRS,and CQ,the proposed network extracts the accuracy of the buildings in the three data sets to 95.6%,98.0%,and 97% respectively;the average merge ratio reaches78.1%,88.5%,83.0%;the recall rate reached 94.6%,98.1%,96.9%,respectively,the accuracy is better than the comparison network U-Net,SegNet,verifying the effectiveness and robustness of the proposed network.Finally,a multi-scale information building extraction network is proposed to realize the automatic extraction of buildings in large-scale remote sensing images.Since traditional networks can only extract single-scale features,the proposed network adds a multi-scale information extraction module between encoding and decoding.The module consists of parallel cavity convolutions.By fusing multiscale features extracted from different convolution branches and using the fusion result as input to the decoding network,the building extraction accuracy is improved.In the test phase,different types of buildings in different areas were used for extraction experiments.The experimental results show that the proposed network can extract the contours of buildings very well.The accuracy of building extraction reaches 97.33%,and the average merge ratio is89.03%.97.25%.