Studies On Object-Oriented Feature Selection And Classification Of High-Resolution Remote Sensing Images

In recent years,with advances of remote sensing technology,and constantly improvement of satellite sensors'spatial resolution,much more image data with abundant and distinct surface information has been provided to researchers and substantial facilitation has been provided to the land cover research.At the same time,to overcome the shortcomings of the traditional pixel-based image analysis mode,such as the phenomena of salt and pepper,same objects with different spectrum and different objects with same spectrum,the idea of OBIA(Object Based Image Analysis)emerged at the right moment.The object here refers to the connected region composed of pixels and objects could usually be obtained through image segmentation process.Compared with pixels,objects own more plentiful characteristics including spectral features(such as visible spectrum,brightness value and normalized differentia vegetation index),shape features(such as length-width ratio,area and shape index),texture features(such as homogeneity,heterogeneity,contrast ratio and angular second moment)and so on.Proper utilization of the above features could improve extraction accuracies of land objects in the pattern recognition.However,with the convenience brought by rich features,curse of dimensionality could be caused and classification accuracies may be reduced because of feature redundancy.Therefore,feature selection is essential for object oriented image analysis of high resolution images.In addition,the commonly used supervised classification mode could not avoid the effect of subjective factor as artificial expertise should be provided before classifications and it would sometimes cause loss of classification performance.So,it is necessary to carry out studies on semi-supervised or even nonsupervised classifications.At present,researches on ensemble classification mainly focus on integration of the same kind of classifiers.To combine advantages of various classifiers,it's essential to study the integration of different kinds of classifiers.Whether in feature selection or classification of the current remote sensing image processing area,machine learning methods are always indispensable.As GA(Genetic Algorithm)and TS(Tabu Search)are typical heuristic searching algorithms,they are usually used in feature selection.SVM is a nonlinearity supervised learning model and it owns unique advantages in solving classification problems.Therefore,theory basis and research situation of a few machine learning methods are first introduced and their problems in feature selection and classification would be analyzed.Then,with the integration of TS,this paper has proposed a novel feature selection method based on GA.Aiming at the weakness of local convergence,a mixing strategy of GA and TS is proposed.TS is integrated into GA,so the particular memory function of TS could be used to improve the local searching capability of GA and to finally obtain the optimal features subset.Experimental results show that the proposed method could improve the searching ability of GA.An object-oriented and self-optimizing classification approach is put forward to classify high-resolution remote sensing images based on semi-supervised thought,FCM(Fuzzy C-Means)algorithm and SVM(Support Vector Machine).To overcome shortcomings of time-wasting and low representativeness from manual selection of training samples in supervised classification,semi-supervised is integrated into FCM.Through this way,training sample set with high representative and strong reliability could be obtained with a few labeled samples,and classification efficiency could also be enhanced.Then,an iterative classification method with self-optimization is designed based on SVM.Classification experiments are carried out to prove that the proposed method could improve classification precision effectively.A new ensemble classification method is proposed as the stacked generalization is improved by iterative calculation of weight values of base classifiers.Classification precision could be increased to a certain degree by optimization of single classifier.However,classifiers own different operating mechanisms and they put up respective advantages and disadvantages during the classification process,classifier integration is a feasible option to complete each other's advantages and to increase classification accuracies.With the stacked generalization model as the integration framework,weight values of each base classifier would be assessed based on their performance during the cross validation course.Then the above weight values are utilized to guide the final integrated process and by this way a new ensemble classifying method is proposed which could adaptively adjust weights of base classifiers.To validate effectiveness and feasibility of the above method,with SVM,BP(Back Propagation),NB(Naive Bayes),C4.5 and KNN(K Nearest Neighbor)as base classifiers and MLR(Multiple Linear Regression)as the meta classifier,the new ensemble classifier is made up.In the final,classification experiments on WorldView-2 and QuickBird images proves the effective and high precision of the proposed method.