Autoencoder (AE) has been widely used in the field of hyperspectral anomaly detection. It is assumed that the background can be reconstructed well, but the anomalies cannot. Hence, the pixels with larger reconstruction error are considered as anomalies. However, owing to the strong nonlinear representation ability of AE, it is difficult to distinguish between background and anomalies. To address this problem, we propose a Residual Self-Attention-based AutoEncoder (RSAAE) for hyperspectral anomaly detection. RSAAE consists of dense residual self-attention modules, an encoder, and a decoder. First, a novel residual self-attention module is designed, which can effectively extract the main features and weaken the ability of subsequent network to reconstruct anomalies, as well as preserve the original features to avoid the deterioration of network performance after the use of dense self-attention modules. Furthermore, inspired by manifold learning, we assume that the background is low-rank in the original space, and has the same property in the latent space after dimensionality reduction. We proposed a low-rank loss function to constrain the latent space, thereby suppressing anomaly reconstruction. Experiments on four real hyperspectral image (HSI) datasets showed that the proposed RSAAE method can produce more accurate detection results than eight popular methods.

The accuracy of hyperspectral target detection (HTD) is often affected by the problems of spectral variation and complex background distribution. Inspired by the powerful representational ability of deep learning, we proposed a 3-D convolution-based global spatial-spectral attention network (GS ² A-Net) to deal with spectral variation in hyperspectral images (HSIs). GS ² A-Net uses 3-D convolution kernels of different sizes to capture local spatial and spectral features to achieve multi-scale information interaction. Different from the previous 2-D attention mechanisms, GS ² A-Net simultaneously considers the information in the spatial and spectral dimensions, and creates a weight map consistent with the size of the original HSI. Furthermore, we proposed a new background suppression strategy based on spectral angle mapping (SAM) to achieve more accurate target detection, which can preserve the targets as much as possible when suppressing the background. The method was validated through experiments on five real-world HSI datasets. Compared with several classical and deep learning-based methods, the proposed method exhibits greater detection accuracy.

To address the problems of noise interference and image blurring in hyperspectral imaging (HSI), this paper proposes a denoising method for HSI based on deep learning and a total variation (TV) prior. The method minimizes the first-order moment distance between the deep prior of a Fast and Flexible Denoising Convolutional Neural Network (FFDNet) and the Enhanced 3D TV (E3DTV) prior, obtaining dual priors that complement and reinforce each other’s advantages. Specifically, the original HSI is initially processed with a random binary sparse observation matrix to achieve a sparse representation. Subsequently, the plug-and-play (PnP) algorithm is employed within the framework of generalized alternating projection (GAP) to denoise the sparsely represented HSI. Experimental results demonstrate that, compared to existing methods, this method shows significant advantages in both quantitative and qualitative assessments, effectively enhancing the quality of HSIs.

Existing deep learning-based hyperspectral anomaly detection methods often overlook frequency domain features, hindering the ability to effectively distinguish between background and anomalies. Furthermore, many methods directly apply Mahalanobis distance to reconstructed hyperspectral image (HSI) for detection, disregarding the structural features of the original HSI. This leads to insufficient representation of important information and ultimately limits detection performance. To resolve these challenges, this paper presents a dual-branch network based on frequency domain learning (DB-FDLNet). Using the Haar wavelet transform, the original HSI is divided into highand low-frequency components. Based on the edge detailed properties of the high-frequency component and the smoothness of the low-frequency component, distinct network branches are designed for feature extraction, with the extracted features fused for HSI reconstruction. Notably, the detection process utilizes the outputs of the high- and low-frequency branches directly, rather than the reconstructed HSI. To fully leverage the key structural information in the original HSI, a Mahalanobis distance detection method incorporating the structural similarity index is proposed. By weighting the covariance matrix, the method enhances critical structural and spectral features, improving detection accuracy while effectively suppressing noise. Experiments on six datasets demonstrate the proposed method's superiority and robustness over eight advanced hyperspectral anomaly detection methods.

Hyperspectral anomaly detection is a crucial technique for recognizing abnormal pixels in hyperspectral images (HSIs), that is, those with distinct spectral characteristics from those of the surrounding background. Traditional methods always fall short in effectively leveraging the information regarding the spectral and spatial aspects of the dataset simultaneously, limiting their detection performances. This article proposes a novel framework using U-Net, termed hybrid convolution and transformer-based U-Net (HCT-Unet), which integrates convolution with a multihead attention mechanism in Transformer for enhanced hyperspectral anomaly detection. To ensure a more comprehensive understanding of spatial and spectral interactions, the HCT-Unet architecture capitalizes on the strengths of local feature extraction of convolutional layers and the capabilities of the long-range dependency modeling of Transformers. A key innovation of this framework is an error attention mechanism, which facilitates adaptive multiscale feature fusion and enhances the feature representation capacity. Furthermore, a new anomaly score calculation method is proposed, which combines reconstruction error with the pixelwise structural similarity index (SSIM) to determine pixel anomaly from both local structural preservation and global spectral consistency perspectives. Experiments carried out on seven different hyperspectral datasets reveal that the proposed method consistently outperforms the widely accepted state-of-the-art methods in hyperspectral anomaly detection.

Spatio-temporal fusion is a common approach in remote sensing, used to create time-series image data with both fine spatial and temporal resolutions. However, geometric registration error, which is a common problem in remote sensing relative to the ground reference, is a particular problem for multi-resolution remote sensing data, especially for images with very different spatial resolutions (e.g., Landsat and MODIS images). Registration error can, thus, have a significant impact on the accuracy of spatio-temporal fusion. To the best of our knowledge, however, almost no effective solutions have been provided to-date to cope with this important issue. This paper demonstrates the robustness to registration error of the existing SParse representation-based Spatio-Temporal reflectance Fusion Model (SPSTFM) [19]. Different to conventional methods that are performed on a per-pixel basis, SPSTFM utilizes image patches as the basic unit. We demonstrate theoretically that the effect of registration error on patch-based methods is smaller than for pixel-based methods. Experimental results show that SPSTFM is highly robust to registration error and is far more accurate under various registration errors relative to pixel-based methods. The advantage is shown to be greater for heterogeneous regions than for homogeneous regions, and is large for the fusion of normalized difference vegetation index (NDVI) data. SPSTFM, thus, offers the remote sensing community a crucial tool to overcome one of the longest standing challenges to the effective fusion of remote sensing image time-series.