Hyperspectral Image Classification On Structured Sparse And Low-rank Representation

The past few years have witnessed tremendous advances in hyperspectral image classification. It is an important task for many practical applications, such as precision agriculture, monitoring and management of the environment, and security and defense issues. With the development of hyperspectral, the resolution of spectrometer can achieve nanoscale. This rich spectral information in hyperspectral image increases the capability to distinguish different physical materials and objects, which leading to the potential of a more accurate hyperspectral image classi?cation.Recent research has demonstrated that sparse representation( SR) is a powerful hyperspectral image representation model. The idea is to represent an input signal as a linear combination of a few items from an over-complete dictionary, which achieves impressive performance on hyperspectral image classification. Recently, low-rank representation( LRR) method has achieved great success in subspace clustering. Comparing to the widely used SR, LRR is better at handling the global structures and correcting the corruptions in data automatically.To this end. we propose a discriminative, structured sparse and low-rank framework for hyperspectral image classification. Firstly, we build a structured sparse and low-rank model, and learn a discriminative dictionary by taking the class labels information of training data into consideration, which makes the learned dictionary robust in the presence of noise and outliers. Secondly, the proposed model can be solved by the alternating direction method of multipliers( ADMM). Finally, to improve the accuracy of imagery classification, we jointly take the spatial and spectral features into consideration. Experimental results on two hyperspectral images show that the proposed algorithm yields higher classification accuracy than the five compared methods( SVM, OMP, OMP-S, SDL and SADL), The overall accuracy can reach 97.75% and 91.82% respectively, improving by 0.42%-22.97%. which demonstrate the efficiency of our proposed algorithm.