Research On Hyperspectral Image Classification Based On Classification Accuracy Prediction

Hyper-spectral images are a combination of image with spectrum. They have rich spectral and spatial information, which make them advantageous to image classification and recognition. However, there exist two challenging problems that affect the classification efficiency and accuracy. First, due to the enormous spectral bands, both the computational burden and redundant information are increased. Moreover, they may cause the curse of dimensionality. Second,’the same spectral from different materials’and’the same material with different spectral’phenomena have a negative effect on the spectrum reliability. To this end, research on band selection and classification methods of hyperspectral images will be conducted to alleviate these problems. The main work and achievements about this thesis are summarized as follows:(1) Establishment of classification accuracy prediction models. Noting that classification accuracy can reflect the separability of bands for classification, they are beneficial for effective band selection and band combination for hyperspectral image classification. To this end, under the hypothesis of Gaussian Mixture Model (GMM), this thesis builds the classification accuracy prediction models by analyzing their denifitions. Since these models simply involve the distribution parameters of each class, they provide the theoretical basis for the judgment that the distribution parameters are the main factors that affect the classification accuracy. Additionally, the explicit prediction models make it available to predict the classification accuracy without training samples. To evaluate the effectiveness of the models, experiments on three hyper-spectral data sets (ROSIS-. RetigaEx, and AVIRIS) and two multi-spectral data sets (GeoEye-1and Z3) are conducted. Experimental results show that the predicted classification performance evaluation measures are highly consistent with the true ones.(2) Overall accuracy prioritization and K-L divergence analysis based unsupervised band selection. Unsupervised band selection methods are usually of low computational complexity. While due to the absence of training samples, it is hard for them to get the criteria that directly reflect the separability of bands. Overall accuracy can depict the separability of bands. Then, based on the overall accuracy prediction model, this thesis first proposes an unsupervised overall accuracy prediction method. To accomplish the comprehensive consideration of separability and redundancy, an overall accuracy prioritization and K-L divergence analysis based unsupervised band selection is developed. Compared with traditional unsupervised band selection method, since this method considers separability of bands, it is more classification-oriented. Experimental results on the ROSIS and RetigaEx data sets show that, it can achieve high classification accuracy.(3) Band selection based on a joint optimization of overall accuracy and redundancy. To get a band subset with maximum separability and minimum redundancy, this thesis formulates a joint maximum overall accuracy and minimum redundancy objective function via a significance weight (SW), which characterizes the importance degree of every band for classification. Thus, band selection is transformed into a constrained optimization problem. In the optimization model, to trade off the overall accuracy and redundancy, an adaptive balanced parameter related to the number of selected bands is designed. Compared with the previous method, this method can achieve better performance at the expense of the complexity of the algorithm. Experimental results on the ROSIS and RetigaEx data sets show that, the proposed method is robust, and outperforms four representative methods in terms of both classification accuracy and redundancy.(4) Weighted spectral-spatial hyper-spectral image classification via class-specific band contribution. The importance of each band in distinguishing different classes is different. Aimed at make full use of each band and improve the usage of spectral information, this thesis introduces the band contributions and formulates a weighted spectral posterior probability model (WSP). To exploit the spatial information, WSP is further combined with the spatial consistency constraint via an adaptive trade-off parameter. In the method, a semi-supervised F-measure prediction method (SS-FP) is developed, which is used to measure class-specific band contribution. SS-FP allows us to estimate the F-measures with the whole data of each band. Thus, the insufficiency and imbalance of the training data are avoided. Experiments on ROSIS, RetigaEx, and AVIRIS datasets show that the proposed method outperforms many state-of-the-art methods.