Research On GF-2 Image Feature Classification Based On Object-oriented And Random Forest Algorithm

How to obtain land cover information efficiently and accurately from high-resolution remote sensing image is a hot topic in the field of remote sensing.Traditional pixel-based target recognition and extraction methods only use spectral features to extract information,ignoring spatial and texture features,which can’t meet the needs of high-resolution remote sensing images to achieve high-precision classification and extraction of ground features.However,the object-oriented interpretation method overcomes the deficiency of pixel classification.This method takes the image object as the basic unit,eliminates the "salt and pepper effect" caused by a single pixel,and at the same time breaks through the limitation of spectral characteristics,and incorporates the characteristics of spatial neighborhood attributes and other auxiliary information into the classification process.As an important machine learning algorithm,random forest,with its advantages of being applicable to multi-dimensional data variables and strong stability,performs well in the classification of remote sensing images,and is widely used in object-oriented interpretation of high-scoring remote sensing images.In this paper,GF-2(Gaofen-2)remote sensing image of a certain area in xinzheng city is used as the data source,and object-oriented technology and random forest algorithm are combined to study the classification of ground objects.The research focuses on three aspects: image segmentation,classification feature space optimization and classifier model construction,aiming at realizing high-efficiency and high-precision classification and extraction of ground feature information from images in the study area.The main research work is as follows:(1)Multi-scale image segmentation based on eCognition software.Under the support of eCognition platform,various parameters that affect image segmentation results are studied: band weight,heterogeneity factor weight and segmentation scale.According to the features and prior knowledge of the ground objects in the study area,the weight of each band of image multi-scale segmentation is determined to be 1:1:1:1,and the weight of shape factor and compactness factor are 0.7 and 0.5.Combined with ESP 2(estimation of scale parameter)tool,the optimal segmentation scale of the image is 170.(2)Research on random forest classification based on feature optimization.In order to improve the efficiency and accuracy of GF-2 image object-oriented random forest classification,a combined feature optimization model based on Relief F and RF-MDA is proposed according to the existing feature selection methods.By optimizing the initial feature space of image objects,the best feature subset is obtained.Based on the best feature subset,the object-oriented random forest classification of the images in the study area is carried out,and the classification accuracy reaches 93.98%,with Kappa coefficient of 0.9250,which is increased by 1.47% and 0.0145 respectively compared with the random forest classification without feature optimization.(3)Comparative study of object-oriented classification algorithms.Based on the same conditions,including image segmentation,sample selection,feature space construction,etc.,different classification algorithms are used to classify the images in the study area.Compare the classification performance of Random Forest(RF)algorithm with K-Nearest Neighbor(KNN),Support Vector Machine(SVM)and CART decision tree(CART)horizontally.The classification results show that the classification accuracy of random forest algorithm is higher than other classification algorithms,and it has high stability in object-oriented feature information extraction of GF-2 remote sensing images.