Research On Classification Of High-Resolution Remote Sensing Imagery Based On Deep Learning

Land cover classification based on remote sensing images is based on the differences between the spectral brightness,spatial structure characteristics and other information of target pixels in different bands in remote sensing images,and is divided into different categories according to certain rules or algorithms.Land use maps,vegetation cover maps and other thematic maps obtained through land cover classification are the basic maps for us to understand the changes of the earth’s surface,for our natural resource management,natural disaster monitoring,geospatial object detection,traffic supervision and urban Planning is important.With the rapid development of my country’s space-to-earth observation technology,the application of high spatial resolution(hereinafter referred to as "high resolution")remote sensing images have become more and more extensive.Aiming at the problem of how to improve the classification accuracy of high-resolution remote sensing images,this paper studies from the data expansion level and the classification algorithm optimization level.The main work and conclusions are as follows:1.First,use the Gaofen-2 remote sensing image as the data source,select three advanced deep learning models U-Net,Seg Net,Deep Lab v3+ and two statistical learning models SVM and RF for classification,and through comparative analysis,three types are obtained.The classification performance of the deep learning model can all be better than the two traditional models,and the classification performance of the U-Net model is better than the two deep learning models of Seg Net and Deep Lab v3+,so U-Net is selected as the basic model for subsequent research.2.Aiming at the problems of complex spatial relationship and lack of spectral information between high-resolution remote sensing images and objects,starting from the data level,by constructing feature images as a supplement to the original images,a multi-feature fusion of high-resolution remote sensing image classification is proposed.method.And using the Gaofen-2 remote sensing image as the data source for verification,the OA value,Kappa coefficient,and F1-score of the proposed method are86.32%,0.84,and 0.85,respectively,which are all better than other comparison models.3.In view of the problem that high-resolution remote sensing images are not effective in classifying multi-scale objects in complex scene classification,starting from the optimization level of classification algorithm,combining the idea of attention and residual learning and the idea of dense connection to the U-Net foundation The structure of the network is further optimized,and a new U-shaped network MSO-UNet with multi-scale output is proposed.It first highlights the main features and suppresses the secondary features by designing a multi-branch attention module and embedding it in each feature extraction layer of the U-Net backbone.Second,a residual-optimized pyramid pooling module is applied to the end of the backbone network to enhance multi-scale feature learning.Finally,the dense connection structure is used at the decoding end to enhance feature reuse,and the previously learned features of different scales are aggregated to further enhance the recognition ability of features and improve the classification accuracy.Classification experiments are performed on the Vaihingen dataset,and the generalization ability of the MSO-UNet network is verified on the Gaofen-2 image.The OA value and Kappa coefficient of MSO-UNet on the Vaihingen dataset are 91.31% and 0.89,respectively,which are better than other comparison models.There are 26 figures,8 tables and 82 references.