Abstract:

Some traditional algorithms applied in hyperspectral remote sensing image classification have some problems such as low computing rate, low accuracy and hard for convergence. According to sparse representation theory, a classification model based on adaptive sparse representation (ASP) is constructed. The algorithm collects a few training samples from a structured dictionary, clusters the error vectors of each step, and signs the cluster center as new atoms making the dictionary. Then the testing samples are regarded as a linear combination of a few training samples of the structured dictionary so as to make the dictionary more suitable for a spare representation of samples. The ASP model is applied to the hyperspectral image of the Washington captured by an HYDICE sensor, and the experimental results show that it has more advantages in the classification in contrast with principal component analysis classifier, linear discriminant analysis classifier, neural network classifier and support vector machine classifier. The overall accuracy of the proposed algorithm is improved by 12% as compared with other methods, which demonstrates the effectiveness of ASP.