Shearlet-Based Region Map Guidance for Improving Hyperspectral Image Classification

Abstract

The inclusion of the spatial context in Hyperspectral Images’ classification tasks has widely proved its efficiency. However, when the neighboring pixels do not represent the same land cover, considering all of them might confuse the classifier and decrease the classification accuracy. To overcome this issue, we propose a Shearlet-based Region Map Joint Sparse Representation (RM-JSR), where the objective is to elaborate a map of edge-surrounded partitions, each having a unique label and each referring to a single land cover. To do so, we first decompose the image using Shearlet Transform. Next, we select the finest scale of the obtained decomposition, where we generally find the salient information about the edges. Then, we carry out a K-means algorithm to segregate the coefficients of the kept scale into edge and not-edge clusters. Afterwards, we apply the Inverse Shearlet Transform and create an image by fusing only the reconstructed edge bands. Finally, we apply a threshold in order to get the region map where homogeneous regions are well delimited. Into the objective function of JSR, we inject the proposed Region Map via Hadamard product. This way, we guide the Simultaneous Orthogonal Matching Pursuit (SOMP), an implementation of the JSR paradigm, in an attempt to overcome its fixed window issue. Compared to other methods attempting to solve this problem, our proposed method achieves better overall classification accuracies.