Spacial-Spectral Classification Of Hyperspectral Remote Sensing Image Based On Ensemble Learning

With the rapid development of remote sensing technology,the spectral resolution of hyperspectral remote sensing image becomes more and more high, and it combines traditional image dimension information with spectral dimension information. Hyperspectral remote sensing image has the characteristics of image and spectral. Although these features make it have obvious advantages in classification, it also has the disadvantages of large amount data and information redundancy. For many conventional classifiers, enough training samples are needed in order to obtain better classification accuracy. Because the cost of artificial marking samples is expensive, it is worth to study algorithm to obtain acceptable classification accuracy in the case of small sample set. This thesis is focused on how to improve the classification accuracy of hyperspectral image based on the existing methods in the case of small training sample set. Ensemble learning, that is, the combination of multiple classifiers, can further improve the performance of single classifier. At present, many scholars have applied ensemble learning for hyperspectral remote sensing image classification. Most of these methods only make use of spectral information, ignoring the rich spatial information of hyperspectral image. This thesis presents several classification methods based on ensemble learning and spatial information. Specific research contents are as follows:(1) A Spatial-spectral classification of hyperspectral remote sensing image based on ensemble learning is proposed. First, the classification result of each weak classifier and segmentation result are combined by majority voting. Then all classification results are combined by majority voting. This method combines ensemble learning and spatial information. It can improve the classification accuracy of each subset. The accuracy of final classification can be improved.(2) A Spatial-spectral hyperspectral remote sensing image classification method based mathematical morphology and ensemble learning is proposed. In the ensemble SVM(Support Vector Machine) learning,the spatial information extracted by using mathematical morphology and spectral information extracted by random spectral feature selection are stacked together. These different data sets are used by each SVM classifier respectively. Then all weak classifier results are combined by majority voting. This stage makes full use of the advantage of spatial information and ensemble learning, which overcomes Hughes phenomenon of the SVM and greatly improves the classification accuracy. Finally the results of ensemble system and image segmentation results are combined by majority voting in order to further improve the classification accuracy. Experiments show that the proposed method obtains an excellent classification performance using small training sample set.