Research On Automatic Building Extraction Method Of High Spatial Resolution Remote Sensing Image Based On Deep Learning

High spatial resolution remote sensing images are widely used in the industry.The pixel-based image classification method and object-oriented image analysis method are limited in the extraction of high spatial resolution remote sensing image information.Deep learning has gradually become a new solution in the field of remote sensing due to the advantages of automatically extracting high-level features of images.This paper aims at the problem of low efficiency and automation in extracting building information from high spatial resolution remote sensing images.Using transfer learning to build a deep learning network,this paper proposes an efficient and rapid extraction of high spatial resolution remote sensing image building information method.The research of the paper is mainly reflected in the following three aspects:(1)Convolutional feature layer analysis of high spatial resolution remote sensing image building extraction.Using the WHU-RS19 high spatial resolution remote sensing image data set,the characteristics of the convolution feature layer for building extraction by convolutional neural network are analyzed.Research shallow features tend to detect image edges,the detected content is comprehensive,and key information is also extracted.As the level deepens,the feature map becomes more abstract and the extracted features decrease.(2)High spatial resolution remote sensing image building extraction model construction.Through experimental comparison and analysis of typical fully connected neural networks Unet,Segnet and Deeplab v3 for building extraction performance,Deeplab v3 was determined as an improved basic network structure.Upsampling Deeplab v3 shallow features to better preserve building edge contour information,the specific approach is: feature extraction of the image through the Res Net 50 network,and the extracted last layer of features through ASPP(void space pyramid pooling)The multi-scale hole convolution of the layer,the features of different scales are fused into the final encoder layer output;the output feature map is restored to a resolution of 4times the size through bilinear interpolation,and then stitched and fused with the feature map of the corresponding scale.Before stitching,1?1 convolution is needed to reduce the number of channels of the feature map,that is,multi-channel merge and fusion;use the pooled index for upsampling,restore twice the size,and stitch and merge again with the corresponding scale features;finally use double The linear difference is restored to the original image size.Through comparative experiments,it is determined that the initial value of the learning rate is set to 0.0002 and the batch size is set to 32 to optimize the performance of the model.(3)High spatial resolution remote sensing image building extraction model training and verification.Using the data set of AI classification and recognition of satellite images in the CCF Big Data and Computational Intelligence Competition(BDCI)hosted by the Chinese Computer Society in 2017,the model was trained to prove the effectiveness of the model.At the same time,the World View-2 high spatial resolution remote sensing image of Jiaozuo City was used for model verification.Experiments show that the model has a certain generalization ability.