Superpixel-based HSI classification via semisupervised K-SVD and multi-scale sparse representation

Hyperspectral image (HSI) classification is one of the most important techniques in HSI-based measurement. In this paper, a semi-supervised k-singular value decomposition (K-SVD) algorithm is proposed, and based on this, a superpixel-based hyperspectral image classification method combining semi-supervised K-SVD and multi-scale sparse representation (SK-MSR) is further proposed. The semi-supervised K-SVD is proposed to solve the problems that the K-SVD algorithm is not good at processing small sample signals and there is no class distinction, which expands the number of training samples by superpixels and uses the joint sparse model (JSM) to solve the sparse problem in the dictionary learning process. After obtaining an over-complete dictionary, in order to effectively use spatial information and remove salt-and-pepper noise, a multi-scale sparse representation superpixel classification algorithm is proposed. Through the performance comparison experiments on two datasets, the superiority of the SK-MSR algorithm relative to other algorithms is demonstrated.