Efficient Graph Convolutional Self-Representation for Band Selection of Hyperspectral Image

Abstract

Hyperspectral image (HSI) band selection (BS) is an important task for HSI dimensionality reduction, whose goal is to select an informative band subset containing less redundancy. However, traditional BS methods basically work in the Euclidean domain and thus often neglect to consider the structural information of spectral bands. In this paper, to make full use of the structural information, a novel BS method termed as Efficient Graph Convolutional Self-Representation (EGCSR) is proposed by incorporating graph convolution into the selfrepresentation model. Since the proposed method is typically modeled in the non-Euclidean domain, it tends to result in a more robust self-representation coefficient matrix. We provide a closed-form solution to the EGCSR model, which leads to high computational efficiency. We further propose two strategies to determine the informative band subset from the coefficient matrix. The first is a ranking-based strategy (EGCSR-R), which ranks every band by calculating the cumulative contribution, and the second is a clustering-based strategy (EGCSR-C), which treats band selection as a band clustering task based on using subspace segmentation. Extensive experimental results on three real HSI datasets show that the proposed EGCSR model is dramatically superior to many existing band selection methods and with high computational efficiency.

SRL-SOA: Self-Representation Learning with Sparse 1D-Operational Autoencoder for Hyperspectral Image Band Selection

“…where diag(A) = 0 is used to prevent trivial solutions that each band is represented by itself. In general, the approaches [3,10] for SRL use the closest norm relaxation of (2) that is 1-norm as it is more efficient and practical to solve.…”

“…The band selection performances have been evaluated based on the classification results obtained by the SVM classifier after applying the band selection procedure. In this manner, the proposed approach is compared against the following band selection methods: SpaBS [1,2], EGCSR [3], ISSC [10], and PCA. In the following, the experimental datasets and settings are first explained and then we present the band selection performances.…”

“…Hyperspectral imaging sensors are able to capture the observed scene with hundreds of different wavelengths. Hence, these optical sensors provide rich spectral information about the target and they have been used in many applications such as target detection [1], land-cover classification [2,3], face recognition [4], and medical imaging [5]. However, such rich spectral information introduces several drawbacks and limitations in hyperspectral image (HSI) data processing related to computational time complexity and memory.…”

“…In Sparse Representation based Band Selection (SpaBS) methods [1,2,9], they set a constraint that the representation of the HSI data is sparse in the transformed domain and corresponding non-sparse coefficients denote only the most descriptive bands. In [3], Efficient Graph Convolutional Self-Representation (EGCSR) is proposed for band selection. Accordingly, they design the EGCSR model using graph convolution where the traditional model of self-representation is extended by considering each band as a node in the non-Euclidean domain over a graph.…”

Hyperspectral Image Classification Based on Sparse Superpixel Graph

“…Then a graph-based semi-supervised HSI classification framework was proposed. Recently, graph convolutional network (GCN) schemes and its various improvements have been proposed to classify HSI [48][49][50][51]. GCN adopts graph to carry out the convolution on arbitrarily structured non-Euclidean data to improve the classification performances of convolutional neural network with fixed size and weight convolution kernel.…”