Neighborhood Activity-Driven Representation for Hyperspectral Imagery Classification

In the classic sparse representation (SR)-based models and their improved versions with the spatial consistency, such as joint representation (JR)-based frameworks, the sparse coefficient is generally considered with the dictionary together for representation. In fact, there is latent significance and property under the sparse coefficient which can be further exploited for classification. In this article, we first introduce two important definitions. One is the activity degree (AD) into the coefficient vector, and the other one is the neighborhood activity degree (NAD) into the coefficient matrix. Through the estimation of AD, we establish a simplified and equivalent model to the classic SR-based classifier called AD-driven representation-based classifier (ADRC). Based on the evaluation of NAD, we propose a novel classifier as an extension to ADRC, named NAD-driven representation-based classifier, including the spatial coherence. The proposed methods take advantages of the sparse idea for effective and concise utilization of individual and overall sparsity. Experimental results on three real hyperspectral datasets demonstrate their efficiency and improvements over the SR-based models and their spatial variants.