RS Image Classification Based On Rough Sets Theory And SVM Classification Algorithm

Remote sensing technology has been widely used in many areas such as:ecological and environmental protection, urban construction, land use, crop yield assessment, resource exploration, atmosphere-ocean monitoring and so on. It also plays an increasingly important role in the socio-economic development. Remote sensing image classification is the premise of the remote sensing technology, and one of the key content in that area. Remote sensing image classification is a typical pattern recognition problem, properties of surface features and classification is the specific application of pattern recognition technology in the field of remote sensing. Because of a wide range of remote sensing image in surface features, image spectral data may still have the synonym spectrum and the same spectrum of foreign body phenomenon. These factors have led to some problems like huge computational complexity, the classification accuracy low, and poor generalization ability in the use of computer classification.It is a very critical problem to choose a classifier in pattern recognition. Ideal classification effect cannot be obtained when choosing some traditional classifier algorithm for remote sensing image classification because of the complexity of the remote sensing image. Support Vector Machine (SVM) classifier, which based on Statistical learning theory, Can solve the actual problem such as the small sample, nonlinear, high dimension. So. SVM has become one of hotspots in the field of pattern recognition research. In this paper, SVM has been used for remote sensing image classification as a result of it's the global optimal, strong adaptability, and promotion ability and high accuracy of classification was obtained. At the same time, through do some experiments to analyze the impact of the SVM four kernel function and its parameter settings on the remote sensing image classification performance.As a kind of supervision classification algorithm, the results of the classification of SVM are closely related to the selection of training samples. Labeling the training samples is overly dependent on the researchers'eyes, which may affect the classification accuracy and efficiency. This paper analyzes the typical samples and atypical sample of the results of the SVM classifier, and find that atypical samples(Mixed pixel) to train SVM classifier has better classification results. And according to the principle of the SVM classification algorithm, we can to explain the reasons for this situation. On this basis, a kind of automatic selection methods of training samples was presented by using the Fuzzy c-Means (FCM) which is unsupervised clustering algorithm. The experimental results show that training SVM classifier for the automatic annotation of the training samples has higher classification accuracy.In addition to the classifier, feature extraction and selection can also affect the classification results. How to find the right characteristics is becoming the key problem in image classification and recognition when the purpose of classification was determined. TM images as the main object of study in this paper, a number of studies have suggested that adding new characteristics to Spectral Signature of TM remote sensing image could improve the classification accuracy. On the basis of the original spectral characteristics, four spectra indexes characteristics which include NDVI, NDWI, MNDWI and NDBI, LBV characteristics, GLCM texture feature, Gabor texture feature and Basic spectral features were used, and then through the experimental comparison classification results in different characteristics. It was found that add new features to make the classification results can improved to varying degrees, but each feature has its own defects. Therefore, by attribute reduction in rough set theory thinking, remote sensing image as a knowledge representation system and then use based on discernibility matrix attribute reduction algorithm on all the characteristics combinations of remote sensing data reduction. It does not depend on a priori knowledge of the case, and reduction of the feature combination and participate in classification to achieve the automation of feature selection. The experiments showed that automatically filter out the characteristics have a good classification results.On the basis of the proposed SVM classifier training samples automatically marked and features automatic screening, which constructed a remote sensing image automatic classification system. And test the system through a complex TM images and QuickBird images, Then compared with several commonly used remote sensing image classification method, the experimental results show that the system of these two images are ideal classification. The platform has the ease of use and scalable, therefore it can be used as the platform of the future scientific research and experimental.