Fast Low-Rank Decomposition Model-Based Hyperspectral Image Classification Method

Abstract

In hyperspectral image classification, jointly using the pixels in an image patch can generally improve the performance. Recently, a new hyperspectral image classification method, which is based on low-rank decomposition model, was proposed by Chen et al. Although this algorithm can achieve state-of-the-art performance and outperform many contemporary classification techniques by jointly classifying both the central pixel and its similar neighboring pixels in the image patch, its computational load is heavy due to its pixelwise processing scheme, which limits its possible applications in practice. In this letter, we proposed a fast groupwise version of this method. In this improved method, the low-rank decomposition model is first used to partition the hyperspectral scene into many groups of similar pixels. Then, the pixels in each similar group would be classified jointly and assigned the same class label. The proposed groupwise classification method can achieve similar performance with the approach described by Chen et al. but with much reduced computational load. Our results demonstrate that when more observations are simultaneously used, the performance of the jointly sparse regression model would increase, and when the number of observations reaches a certain amount, the performance would no longer increase significantly. Our experimental results also demonstrate that, when the jointly sparse regression model is used, the quality is preferred to the quantity of the observations after the number of observations has reached a certain amount.