Research On Object-oriented Multi-level Classification Method Of GF-2 Remote Sensing Image

With the continuous breakthroughs in earth observation technology,high-resolution remote sensing images have been continuously developed and enriched.Highresolution remote sensing images containing a lot of detailed information are used as the data basis,which makes object-oriented image analysis methods widely used in the field of high-resolution remote sensing image processing,and gradually becomes one of the main methods of high-resolution image processing.Compared with traditional pixel-based classification,object-oriented classification can use the rich spectrum,shape and texture information of the image to a greater extent to avoid possible "salt and pepper",but object-oriented classification based on a single scale is difficult to fully Extracting detailed information of various types of features,so multi-scale objectoriented multi-level classification has become a research hotspot.In the past,multi-level classification methods generally use threshold method or membership function to establish classification rules in information extraction.This article takes GF-2 image as an example,selects some areas of Beihu in Changchun City as the research area,expounds the key technology of object-oriented classification,and proposes an object-oriented multi-level classification method based on classifier: First,the ESP scale evaluation algorithm is used to calculate the most The possible values of optimal segmentation scales are used to visually discriminate and select the optimal scales of local features,and establish a multi-level structure according to the optimal scales of local features;secondly,use recursive feature elimination method for preliminary optimization to eliminate redundant features in the feature space;Use correlation matrix to calculate the correlation between features,remove the features with high correlation and low importance ranking,realize the second optimization,and ensure the independence of each feature;Finally,use the classifier to extract the feature information of each layer 3.Realize the inheritance between layers and get the final classification results;and make a comparative analysis with the traditional single-level classification method.In this paper,the single-level classification method uses the segmentation quality function to quantitatively calculate the global optimal segmentation scale,thereby establishing a single-level structure,and using the optimized feature space and classifier for single-level classification.In order to avoid the accuracy deviation caused by the classifier,the experiment uses the CART decision tree algorithm and the random forest algorithm with good effect in land classification to perform single-level classification and multi-level classification,and finally evaluates the accuracy of the four classification results.The research results are as follows:1)Calculation of optimal segmentation scale.Quantitatively calculate the global optimal segmentation scale using the segmentation quality function;use the ESP(Estimation of Scale Parameter)scale evaluation algorithm to calculate the possible values of the optimal scale,and determine the surface covered buildings,vegetation,water bodies,The optimal segmentation scale of bare land,roads and shadows is used to establish a multi-level structure as the basis for subsequent information extraction.2)The optimization method of feature space.Select the commonly used features in the land use classification to construct the initial feature space,use the recursive feature elimination method for preliminary optimization,remove the features such as Border index,Asymmetry,Area,Elliptic fit,Roundness,Length / Width;use correlation matrix to calculate the features Correlation,remove features with high correlation and low importance,such as NDVI,NDWI,SAVI,Ratio Green,Ratio Red,to achieve the purpose of secondary optimization,and finally get the optimized feature space.3)Object-oriented multi-level classification.In order to compare the classification accuracy of different algorithms,this paper uses CART decision tree algorithm and random forest algorithm for classification.Experiments show that the overall accuracy of single-level classification based on CART decision tree is 0.7001,Kappa coefficient is 0.6497,the overall accuracy of single-level classification based on random forest is 0.7346,and Kappa coefficient is 0.6823;The Kappa coefficient is 0.7543,the overall accuracy of the multi-level classification based on random forest is 0.8586,and the Kappa coefficient is 0.8304.In land use classification,regardless of single-level classification or multi-level classification,the classification accuracy based on random forest algorithm is higher than CART decision tree algorithm.The object-oriented multi-level classification method proposed in this paper implements automatic classification based on classifier classification instead of rulebased classification.Compared with the traditional single-level classification method,this method is better for land use classification accuracy and effect,and has theoretical significance and practical value.