Research On Land Cover Classification Algorithm Based On Multi-source Remote Sensing

Guilin is a national sustainable development agenda innovation demonstration zone approved by the State Council of my country in 2018.Land cover information is the basic geographic data.Mastering and utilizing land cover information is of great significance to promoting the construction of innovative demonstration areas of the National Sustainable Development Agenda and my country’s sustainable development.The domestic high-resolution series and foreign multi-spectral remote sensing images have been widely used in many fields,and also provided a new opportunity for the classification and utilization of land cover in Guilin.Although the land cover classification technology based on remote sensing images has been studied for a long time,how to establish a fast,efficient and accurate land cover classification model is still a challenge for the application research of remote sensing images.Based on the optimization of the U-Net network model,this thesis constructs a network model that combines multi-source remote sensing data for remote sensing image classification.The research content and results of the thesis are as follows:(1)Aiming at the problem that a single remote sensing data source can obtain limited features of ground object information,which is not conducive to the classification of complex ground objects.In this thesis,the multi-spectral images of the Landsat8 remote sensing satellite and the high-resolution images of the Gaofen-2 satellite are combined with the Digital Elevation Model data for image fusion to obtain more ground feature information features.The results show that using Gaofen-2 remote sensing images alone for land cover classification,the overall classification accuracy is relatively low,and there are mixed and misclassified phenomena among various ground objects.When multi-source remote sensing image data is used,more comprehensive features of ground object information can be obtained,the overall classification accuracy has been greatly improved,and various ground objects can be better distinguished.(2)In view of the complex and redundant background interference of remote sensing images,it is difficult to extract useful information features.This thesis proposes a multi-source remote sensing image land cover classification method that improves the U-Net network model.Firstly,the number of convolutional layers is increased in the U-Net network model,which solves the problem that it is difficult to extract the deep features of remote sensing images with complex ground objects due to the shallow level of the U-Net network model;Secondly,add Dropout and batch normalization layers to avoid overfitting to a certain extent,and use depthwise separable convolution instead of standard convolution to reduce the amount of network parameters and improve model training speed and robustness;Finally,the spatial and channel attention fusion mechanism module is introduced to enhance the useful features of the image and suppress the useless features of the image.The results show that the improved U-Net network model algorithm can effectively extract the information features of remote sensing land cover,and obtain land cover classification results with higher overall accuracy and Kappa.(3)In this thesis,the traditional classification algorithm and the deep learning classification algorithm are used to analyze and compare the accuracy of the classification results of land cover.Experiments on multi-source remote sensing image datasets show that the overall accuracy of the deep learning algorithm in land cover classification has been greatly improved.The improved U-Net network model has the highest overall accuracy,followed by the U-Net network model,NN,SVM,FCN＿8S model,and the worst is MLC.The improved U-Net network model is higher than U-Net network model,neural network,support vector machine,FCN＿8S network model and maximum likelihood classification increased by 3.44%,3.93%,4.17%,4.39% and 5.15% respectively.Moreover,the result map of the improved U-Net network model classification method is closer to the real situation of ground objects.