Research On Building Extraction Method Of High-resolution Remote Sensing Image Based On Deep Convolution Neural Network

With the rapid development of the city,many people come into the city to settle down and live,the number of urban buildings has mushroomed.It is necessary to plan urban land resources reasonably,curb the phenomenon of illegal and disorderly construction of urban buildings.Accurate building extraction for high-resolution remote sensing images plays an important role in urban planning,urban management,and change detection of urban buildings.The traditional building extraction method mainly depends on handcrafted features such as texture features,shadow,local structure(edges,lines,and corners),which has achieved some results,but it is time-consuming and laborious.In recent years,with the improvement of computer performance and the development of deep learning,people gradually apply deep learning to the field of remote sensing,which has made some achievements in image classification,target detection,semantic segmentation,and so on.Based on the ability of self-learning multi-level features of the convolutional neural network,this paper proposes a novel network model named FE-Net(Feature Enhancement Network),which has “Encoder-Feature Enhancement-Decoder” structure.This network model can realize end-to-end automatic building extraction for high-resolution remote sensing.The specific research is as follows:(1)Using the Massachusetts building dataset as the experimental data,based on the U-Net model,the network model layers are explored.This article mainly explores U-shaped network models(U-Net5,U-Net6,U-Net7)with five,six,and seven network layers.When the number of network layers is eight,it is restricted by experimental conditions,so we do not consider it.Compare the building extraction effect and accuracy of different network-layer models(U-Net5,U-Net6,U-Net7)in order to find the best basic network model.(2)Based on the best basic network model found in(1),the performance of the network model algorithm is improved by adding feature enhancement structures to the structure of the basic network model.The basic idea is that because there is a continuous down-sampling part in the basic network of the "encoder-decoder" structure,this part is easy to lose detailed information.In order to reduce the loss of detailed information and retain detailed information,a feature enhancement structure based on dilated convolution is added to the structure of the best basic network.The structure can obtain multi-scale features of the image,retain detailed information,and there will be no reduction in resolution.It connects multi-scale through series and parallel connections.The feature maps are added to achieve feature enhancement.Based on the feature-enhanced network model,the Massachusetts Building Dataset is used for building extraction.(3)In order to improve the performance of the network model further based on the model built in(2),it replaced the activation function.Because the activation function ReLU of the above network model is prone to cause the phenomenon of neuron "death" and the weight cannot be updated,and the ELU activation function not only has all the advantages of ReLU,but also can effectively prevent neuron "death" and make up for the shortage of ReLU.Therefore,the activation function ELU is used to replace Re LU,and the network model called FE-Net with the activation function ELU is obtained.Based on the model,building extraction is performed on the Massachusetts building dataset.(4)Using high-resolution remote sensing images—the domestic Wuhan building dataset,for the above network models:(1)U-Net5,(2)U-Net6,(3)U-Net7,(4)“encoder-feature enhancement-decoder” network structure model,(5)“encoder-feature enhancement-decoder” network model FE-Net with the activation function ELU for further verification,and the extraction effect of the building was qualitatively and quantitatively evaluated.Through the above experiments,combined with the extraction results of two sets of high-resolution remote sensing image building datasets in Massachusetts and Wuhan,it can be found through qualitative and quantitative analysis that the proposed network model called FE-Net can better achieve building extraction of high-resolution remote sensing images,and the values of Relaxed F1-measure are 97.23% and 90.12%,respectively.