Research On High-resolution Image Classification Method Based On Multi-feature Parameters

With the continuous breakthrough of remote sensing technology,the spatial resolution of remote sensing images has been improved from tens of meters to sub-meter level.The spectral,texture,geometry and other feature information of high-resolution images are very rich.Using a variety of feature parameters to classify remote sensing images has been become a research hotspot in recent years.Image classification mainly includes two types: pixel-based and object-oriented.The pixel-based classification method is mainly based on the spectral characteristics of the image,which is difficult to meet the classification requirements of high-resolution remote sensing images;Object classification is based on the segmentation of the object as the basic unit and fully combines various feature information of remote sensing images,which can effectively avoid the occurrence of "same object with different spectrum,same spectrum foreign object" and other phenomena.Based on GF-2 remote sensing images,this paper selects the Beihu area of Changchun City as the research area to study object-oriented single-level and multi-level classification and pixel-based SVM and CART decision tree classification methods for high-resolution remote sensing images.First,the multi-level and global optimal segmentation scale is determined by using the ESP scale evaluation tool and the comprehensive heterogeneity index is calculated;secondly,three features of spectrum,texture and thematic index are selected to construct the initial feature space,and the random forest algorithm is used to construct the initial feature space.optimization work;finally,the object-oriented and pixel-based classification methods are used to classify and identify six kinds of ground objects in the study area,and the classification methods are compared and analyzed.The main researches are as follows:(1)Selection of the optimal segmentation scale.Firstly,the optimal band combination and band weight are determined by calculating the OIF index,and the homogenization factor and the optimal scale interval are determined by the method of controlling variables;secondly,the possible value of the multi-level optimal segmentation scale is determined by the ESP scale evaluation tool,and the visual discrimination method was used to determine the optimal segmentation scale at different levels in turn;finally,the comprehensive heterogeneity index was calculated to determine the global optimal segmentation scale.(2)Construction and optimization of feature space.Firstly,J-M distance analysis method is used to evaluate the separability of training samples;secondly,the more commonly used features in image classification are selected to construct an initial feature space,which includes 8 spectral features,6 thematic index features and 16 texture features Finally,the random forest algorithm is used to analyze the importance of the above 30 features and complete the feature optimization,and finally determine the 10 features as the optimal feature combination.(3)High-scoring remote sensing image classification based on multi-feature parameters.In this paper,object-oriented single-level and multi-level classification,pixel-based SVM and CART decision tree classification are used to study image classification methods.The research shows that among the four classification methods,the multi-level threshold classification accuracy is the highest,the overall accuracy is0.8780,and the Kappa coefficient is 0.8420;in the pixel-based classification method,the classification accuracy obtained by SVM is slightly higher than that of the CART decision tree,and the overall accuracy is 0.7940,and the Kappa coefficient is 0.7370;in this paper,whether it is a multi-level or single-level classification method,the accuracy of object-oriented classification is higher than that of pixel-based classification.