Research On The Nearest Regularized Subspace Classification For Hyperspectral Image

With the continuous development of hyperspectral image technology, we can easily obtain the earth surface information. Hyperspectral image classification technology as an important part of the image processing is greatly concerned. The nearest regularized subspace image classification, it combines the collaborative classification with Tikhonov regularization factor for hyperspectral data classification decision, which make it more suitable for the training data.We found that there are some obvious deficiencies in the nearest regularized subspace classification through the study. The Tikhonov regularization based on the Euclidean distance can't represent the difference from the spectral structure effectively. At the same time, this model only takes the spectral characteristics of hyperspectral data into account and waste the spatial information. Therefore, we propose three improvements from these drawbacks.First, the similarity measurement on Euclidean distance can't present the relation of different spectral vectors, especially in same substance. To introduce the additional spectral similarity measurement, such as the spectral angle measurement, the spectral information divergence and so on, improve the performance of the algorithm.Second, given the current classification model only takes the spectral characteristics of hyperspectral image into account and ignore the importance of the structural information, we propose the classification combine the spectral and spatial information, apply the Markov random field model into the nearest regularized subspace model, which classify the different substances of the whole image accurately.Third, based on the domain system, the center pixel is achieved by the surrounding pixels and reduces the impact of irrelevant pixel adaptively. In this paper, we adopt the Gaussian model to render the expression of spatial information. Improve the accuracy of classification algorithm. To give a direction for the spatial information applied in hyperspectral image classification.