Improving hyperspectral data classification of satellite imagery by using a sparse based new model with learning dictionary

Statistic classification of hyperspectral data is a great challenge because of its large number of spectral channels, especially when the labeled training samples are relatively few. Most of the classification methods require using a large number of training samples, but in remote sensing situations, identifying and labeling samples are extremely difficult and expensive. A sparse representation classification approach (SR) has proven that it can perform quite well with only a few labeled samples. In this paper we propose an improved classification model by using dictionary learning of sparse representation (DLSR). We tested the proposed approach with 75%, 50% 10% and 5% amount of training samples and compared the classification accuracy of hyperspectral data with the state-of-the-art SVM, HSVM and the classical sparse representation methods. We performed experiments using real hyperspectral dataset of the NASA AVIRIS spectrometer acquired data over the KSC, Florida on March 23, 1996. Results show that our improved approach offers more classification accuracy and more efficiency than the three above methods.