Research On Sample Optimization Algorithm For Hyperspectral Image Classification

Hyperspectral data has hundreds of dimensions of continuous narrow band,which contains rich spectral information,providing a possibility for the classification and recognition of fine ground object information.With the increase of its application,hyperspectral image classification has become an important research branch in the field of hyperspectral.The classification of hyperspectral images mainly depends on the spectral and spatial characteristics of pixels.The current algorithm combines two features to improve the classification effect in varying degrees.However,there are existence of a large number of mixed pixels in hyperspectral data.On the one hand,the same type of ground object is mixed with other ground objects,or mixed in different proportions,resulting in the increase of spectral differences within the class.On the other hand,target mixing among different classes reduces the spectral difference between classes.Due to the existence of intra-class difference and inter-class coupling,inter-class separability is reduced,thus increasing the difficulty of classification greatly.Therefore,in this thesis,the classification problem caused by mixed pixels is mainly solved from the perspective of sample space optimization.The main work of this thesis includes the following three aspects:(1)For the problem of intra-class spectral differences,this thesis proposes the sample optimization strategy of intra-class reclustering.The samples of the same class were reclustered and clustered into multiple subclasses.When samples were selected,it was ensured that the training samples of each class covered all subclasses,that is,all situations in the class were covered.From the perspective of training samples,it provides a comprehensive data basis for classifier training.(2)For the inter-class spectral similarity problem,this thesis proposes a inter-class sample optimization strategy based on the global clustering of super-pixel blocks.In other words,the whole image is unsupervised clustering with super pixels as the unit,and the clustering results and classification labels are compared to find the region with high similarity between classes.The representative samples of all kinds in this region are selected to improve the classification between classes.Meanwhile,samples are selected in the region within the class according to the sample differences.From the perspective of training samples,the classifier should pay more attention to these difficult samples which difficult to separate in the learning process.(3)For the classification strategy,this thesis uses the integrated learning idea for reference,and uses spatial information to improve the classification results.Firstly,the high confidence region and low confidence region are obtained by combining the classification results of the same scene with two simple classifiers.Secondly,by using the assumption of local spatial consistency and the edge protection filter,the label information of the neighborhood high-confidence region is transmitted to the low-confidence region to achieve the effect of label correction,thus improving the classification accuracy.