Semi-Supervised Classification For Hyperspectral Remote Sensing Image Using Neighborhood Information

Remote sensing image classification is an important technique for land-use and land-cover and environmental monitoring information extraction,etc.with the rapid development of remote sensing technology,mass remote sensing image data can be obtained,it is very important to labeled image,especially to labeled hyperspectral remote sensing image.Thus,semi-supervised classication technology has attracted the hot concern,while many traditional semi-supervised algorithms just only use the spectral information,the classification accuracies are usually decreased by the mixed pixels,ambiguous boundaries,the same spectrum with the different objects and the same objects with the different spectrum,etc.The previous studies have shown that combing spectral and spatial information can improve the classification accuracy in a certain.Inspired by these studies,including the spatial neighborhood information from pixel level and object level,respectively,some spatial-spectral semi-supervised classification algorthims are proposed in the thesis.The details are as follows:1)To enhance the confidence of the candidate samples when choosing samples,a new sample selection strategy is proposed using the multi-classifier integration and spatial neighborhood information,in which the high confidence samples are automatically added to the model,and the semi-spervised algorithm of hyperspectral remote sensing image integrating multiple calssifiers and active learning is constructed.Experimental results show that the proposed method has good stable performance and can improve the classification accuracy with small labeled samples.2)Based on superpixel segmentation,the superpixel is used as the co-opted labeled sample unit instead of pixel.In which the homogeneity of superpixel and its spatial neighborhood information are used to build a semi-supervised classification algorithm of hyperspectral remote sensing image.The experimental results show that the proposed algorithm can reduce the number of iterations and enhance the performance of classification model.3)Because the optimal segmentation scale is difficult to be obtained in the traditional objects region adjacency graph,there are some problems such as inaccurate spectral homogeneity in each object,etc.,which will result in poor application of the merging methods for the traditional objects region adjacency graph.To overcome the limitation,the superpixel based adjacency graph is proposed for co-opting samples.First,the superpixel objects are labeled by using spatial overlay of the original labeled pixels and superpixels.Then we build the region adjacency graph using the labeled superpixel objects as the starting points.Bhattacharyya coefficient beteween superpixels is calculated as the merging criterion of adjacency graph.The last merged superpixes are selected as the initial labeled samples,and the training samples are selected randomly from the initial samples for the model training.To furherly enhance the performance of the classification model,the edge-presering filtering is used to refine the classification map.The experimental results show that the proposed algorithm can obtain good results without multiple iterations.