Low-Rank and Sparse Matrix Decomposition with Cluster Weighting for Hyperspectral Anomaly Detection

Abstract

Hyperspectral anomaly detection plays an important role in the field of remote sensing. It provides a way to distinguish interested targets from the background without any prior knowledge. The majority of pixels in the hyperspectral dataset belong to the background, and they can be well represented by several endmembers, so the background has a low-rank property. Anomalous targets usually account for a tiny part of the dataset, and they are considered to have a sparse property. Recently, the low-rank and sparse matrix decomposition (LRaSMD) technique has drawn great attention as a method for solving anomaly detection problems. In this letter, a new anomaly detection method based on LRaSMD and cluster weighting is proposed. We concentrate on the sparse part, which contains most of anomaly information, and calculate the initial anomaly matrix based on this part. To suppress background regions and discriminate anomalies from the background more distinctly, a weighting strategy in terms of the clustering result is used, and then the anomaly matrix is updated. The judgement of anomalies is made according to the responses on the matrix. Our proposed method considers the characteristics of anomalies from the spectral dimension and the spatial distribution simultaneously. Experiments on three hyperspectral datasets demonstrate the outstanding performance of the proposed method.

Orthogonal Subspace Projection Target Detector for Hyperspectral Anomaly Detection

“…For example, window-based local anomaly detectors, which implemented K or R using local windows [6][7], sliding windows [8], dual windows [9][10], multiple windows [11] and kernel anomaly detector [12], anomaly detection for unlabeled classification [13], real time processing of anomaly detection [14], guided filtering-based AD [15], spectral-spatial feature extraction-based AD [16], background separation-based AD [17], sparsity scoreestimation framework for AD [18]. Most recently, other approaches have been also developed such as deep learningbased anomaly detector, low rank and sparse matrix decomposition (LRaSMD) model-based anomaly detectors [19][20][21][22][23][24], low-rank and sparse representation [25][26][27], autoencoder [28][29][30], generative adversarial network (GAN) [31][32], game theory-based AD [33]. All of these works did not go beyond the original idea of RX/R-AD.…”

Hyperspectral Anomaly Detection Through Sparse Representation With Tensor Decomposition-Based Dictionary Construction and Adaptive Weighting

“…Adaptive weighting can improve the performance of AD by assigning different weights to different pixels depending on the characteristics of each pixel. For example, [38] designs an eight-connected domain division-based weighting strategy to suppress the background signatures in the sparse component extracted by LRaSMD. The principle is that the more the number of pixels counted in a homogeneous region obtained by the eight-connected domain division, the more likely it is that this region belongs to the background.…”

“…The principle is that the more the number of pixels counted in a homogeneous region obtained by the eight-connected domain division, the more likely it is that this region belongs to the background. [38] weights the detection result by analyzing the spatial background aggregation in HSI, while our algorithm performs weighting by mining the anomaly information contained in the anomaly part extracted by Tucker decomposition.…”

Hyperspectral Anomaly Detection via Dictionary Construction-Based Low-Rank Representation and Adaptive Weighting

“…However, the sparse matrix obtained is always contaminated by isolated noise, thus causing some false alarm points [25]. As an improvement, low-rank and sparse matrix decomposition (LRaSMD) [26] extracts noise from the valuable signals, and then further separates the low-rank background and sparse anomalies. The anomaly detector in [27] first extracts some source components by using the unmixing operation, and then identifies the components that are sparse and have the largest accumulated distance from other components.…”