Spectral-Spatial Classification Based on Affinity Scoring for Hyperspectral Imagery

Recently, spectral-spatial classification for hyperspectral imagery (HSI) has become popular since it addresses the issues of limited prior knowledge and spectral internal-class variability. To provide simple and effective approaches in this area, we propose a novel supervised spectral-spatial measurement, affinity score (AS). It considers three factors: local spatial consistency, spectral similarity, and prior knowledge. It is used for classification since it can be directly designed to quantify how much a pixel belongs to a class. Furthermore, we propose two AS-based spectral-spatial classification methods such as combinational rule based on AS (CRAS) and semisupervised classifier based on AS (SCAS). CRAS creates a classification map with increased accuracy by combining spectral classification and spatial segmentation results. SCAS classifies the original HSI in a semisupervised manner. Between the two methods, SCAS is robust to the scarcity of training samples while CRAS is efficient. Experimental results show that the proposed methods can outperform several classic classifiers and state-of-the-art methods.