Weighted Kernel joint sparse representation for hyperspectral image classification

Kernel Joint Sparse Representation Based on Self-Paced Learning for Hyperspectral Image Classification

“…JSR pursues a joint representation of spatial neighboring pixels in a linear and sparse representation framework. If neighboring pixels are similar, making a joint representation of each neighboring pixel can improve the reliability of sparse support estimation [17,19]. The success of the JSR model mainly lies in the follows two factors: (1) joint representation: the neighborhood pixel set is consistent, that is, pixels in a spatial neighborhood are highly similar or belong to the same class; (2) linear representation: the linear representation framework in the JSR model is coincident with the hyperspectral data characteristics.…”

“…[23]. To cope with the nonlinear problem, kernel-based JSR (KJSR) methods are proposed [19,[24][25][26]. KJSR mainly includes two steps: projecting the original data into high-dimensional feature space using a nonlinear map and then performing JSR in the feature space.…”

Guest Editorial: Hyperspectral Imaging and Applications

“…Shamsolmoali et al propose a CNN in network pipeline to reduce the dimension [9], as well as extract spectral‐spatial features for HSI classification. Hu et al improve kernel joint sparse presentation methods by proposing two weighted schemes to incorporate spatial neighbourhood information for HSI classification [10]. Huang et al present an approach for HSI clustering, where sparse dictionary learning and anchored regression are combined to reduce the computational time and improve clustering performances [11].…”

“…The first is the deep learning models, where CNNs have been particularly applied for classification of HSI [8, 9], enriching the previous work in [22] using autoencoders and in [23] using deep belief network. Meanwhile, the sparse representation and sparse learning have also be emphasised in data classification [10, 11], using kernel joint sparse representation and sparse dictionary learning, following the success in [24, 25]. Another trend is the combination of various optimisation algorithms to solve the optimisation problems, such as gravitational search algorithm [13] and Jaya optimisation [16] for band selection and genetic algorithm for target detection [18].…”

“…Not surprisingly, it continues the trend of data fusion in HSI, including spectral‐spatial fusion [8–10, 12, 16, 17] and multiple data sources fusion [21]. It is worth noting the saliency based approach in [17], where spatial saliency and spectral saliency are combined for effective target detection, is an extension of conventional saliency detection in computer vision [29, 30].…”