Remote Sensing Image Classification Based On Multi-featuras Fusion

The classification of remote sensing image has always been an important content of remote sensing research. One of the key questions of remote sensing image research is to deal with multi-class remote sensing image classification and satisfy a certain precision.Through computer numerical processing of remote sensing image, the compute classification of remote sensing image achieves the object goal of automatic classification and recognition. Due to existence of the same object with different spectra and foreign body with the same spectrum, computer remote sensing image classification precise has been influenced greatly. Advancing the present classification method is of great importance to extract surface features precisely, supervise country resource and plan land-use.The thesis takes the TM remote sensing data of urban district of Shi Jiazhuang as the research object. It based on the texture characteristics, spectral features and spatial features to build multi-feature image set. Besides, based on feature fusion remote sensing image classification method, it analyzes the research region’s experiment and precise by using Maximum Llikelihood, Support Vector Machine and decision tree classification. In the first place, the thesis compares and analyzes the texture feature extracting method, which chooses the variance of gray level co-occurrence matrix method and Gist of Gabor filter characteristics and utilizes optimal index method to compare the two. In consequence, it shows that the texture feature performance of Gist is superior than the traditional performance. In the second place, based on remote sensing image to extract the vegetation index(NDVI), Tasseled Cap components and principal component weight which are reflected by the characteristics of the spectrum, it introduces digital elevation model to express the space characteristics. According to the characteristics of the properties of random combination, and analyzes each feature combination classification samples under J-M distance, build the optimal classification feature combination. Finally, multi-feature fusion of remote sensing image has carried on the maximum likelihood and SVM classification which are based on the classification characteristics and extracted fromstratified features. According to the precision analysis, classification characteristics play a positive role on the remote sensing classification accuracy improvement and the original image is improved based on the multi-feature image classification accuracy. The experimental results of three kinds of classification method show that decision tree classification is the optimal decision, accuracy of 87.33%, kappa coefficient for 0.8047;SVM method is followed, accuracy of 85.33%, kappa coefficient of 0.8047; finally, the maximum likelihood method is the last, accuracy of 80.33%, kappa coefficient of 0.7340.