Research On Land Cover Classification Method Of GF-2 Remote Sensing Image Based On Multi-feature Fusion

Land resources have always been an important resource for human survival and development.The rapid acquisition of land resource information based on remote sensing of land use/cover classification has always been the cornerstone of resources and environmental monitoring,territorial and spatial planning,land resource management,etc.It is of great significance to the sustainable use of land resources.With the rapid development of satellite remote sensing technology,the types and quantity of remote sensing data are constantly increasing.Among them,high spatial resolution remote sensing images have richer shapes,textures,spatial information,and more sufficient semantic information than traditional images.They can achieve more accurate extraction of surface information and become one of the important data sources for earth observation and research.At the same time,due to the substantial increase in spatial resolution,the details of ground features have been enlarged,the performance of image features has become more complex,and the difficulty of classification has risen sharply,even leading to a decline in the accuracy of remote sensing interpretation of land cover.In this regard,this paper introduces a deep neural network with strong feature learning ability—U-Net network to achieve a more efficient and accurate land cover classification.However,due to the highly abstract nature of depth model,there is an obvious problem of loss of detailed information,and the high dependence on sample data at the same time limits the application of deep networks in land cover classification.The low-level visual features based on remote sensing images have the advantages of simplicity,strong interpretability,rich detailed information and no dependence on data,but it is difficult to fully explore the deep-level features in the image,and the improvement of classification accuracy is limited.In view of the above problems,this paper takes GF-2 remote sensing image as the main data source,proposes a land cover classification method based on multi feature fusion,realizes the complementary advantages of deep learning and bottom features,and provides a certain reference for efficient and accurate extraction of land cover classification information from GF-2 remote sensing image.The main research contents and achievements are as follows:(1)The feature space of high-resolution image for land cover is constructed,and a feature optimization method integrating Pearson correlation coefficient and average accuracy is proposed.Based on the data characteristics of high resolution remote sensing image and the characteristics of different objects in the image,this paper constructs the original feature space including spectral features,texture features,spatial domain features and terrain features.At the same time,in order to avoid the degradation of classification performance caused by "dimension disaster",this paper proposes a feature selection method integrating Pearson correlation coefficient and average accuracy reduction to obtain the final selected feature subset,and obtains the importance ranking results of spectral features > terrain features > spatial domain features > texture features in the study area.Finally,a comparative classification experiment is carried out with the help of the classical remote sensing classification algorithm maximum likelihood method to verify the effectiveness of the feature selection method.(2)The land cover classification model of U-Net high-resolution image considering the optimal features is constructed.This paper selects U-Net neural network as the basic network,optimizes the u-net model by adding dropout layer and BN layer,data enhancement and expansion of sample set and verification set division,and takes multi band image with optimal features as input data,makes full use of the advantages of u-net depth feature extraction,improves the recognition accuracy of ground objects in complex image background,and successfully constructs the optimal features The land cover classification model of high resolution image based on u-net network.In this paper,through the land cover classification experiment in the study area,the feasibility is verified,and the overall classification accuracy of 83.76% is obtained,with kappa coefficient of 0.7664.The results show that the method has good classification effect,and has certain advantages in the application of land cover classification of high-resolution remote sensing images.(3)A land cover classification method based on multi feature fusion is proposed.In this paper,the serial fusion method is used to fuse the deep semantic features and the bottom visual features and input them into the classifier to complete the classification.A land cover classification method based on multi feature fusion is proposed.In this paper,8 groups of comparative classification experiments are set up,and the random forest classifier is the best classifier,and the depth feature + optimal feature is the best fusion scheme.The overall accuracy of the classification results is 89.64%,and the kappa value is 0.8482.At the same time,compared with the classification results of optimal feature + Maximum Likelihood,optimal feature + Random Forest and U-Net,the accuracy is improved by 23.72%,8.29% and 5.88% respectively.The results show that the classification method based on multi feature fusion has certain advantages for land cover classification of high-resolution remote sensing images,realizes the complementary advantages between different features,and optimizes the classification results.