Hyperspectral image classification (HSIC) has advanced significantly in recent years, driven by the development of advanced algorithms in remote sensing. However, the high-dimensional nature of hyperspectral data and the limited availability of labeled samples remain significant challenges, hindering the effectiveness of many existing methods. To address these limitations, we propose SpiralMamba, a novel classification framework inspired by the recent Mamba model, renowned for its efficient global feature extraction with linear complexity. To minimize the loss of spatial information when converting images into sequences for Mamba processing, we propose the innovative spiral scan embedding (SSE) module. In addition, the introduction of the Gaussian mask weighting (GMW) module enhances the feature weights around the central pixel, thereby improving the classifiability of the extracted features. We introduce the lightweight Mamba module (LWM), which reduces model parameters and computational requirements, making it particularly well-suited for HSIC with limited samples. Experimental results on three real datasets demonstrate that the SpiralMamba model outperforms existing methods in various performance metrics.

Hyperspectral image (HSI) compression presents the challenge of preserving both spectral and spatial fidelity while achieving high compression rates. Current compression methods frequently depend on band-by-band compression or simplistic joint modeling, which complicates the balance between spectral consistency and perceptual quality. To address this issue, a compression driven generation framework (BRC-CDM) is proposed, which decouples the extraction of compressed representations from the reconstruction of high-quality images. We introduce reference band information through channel-level concatenation to guide the spectral residual compression network in collaboratively extracting residual information in both spatial and spectral domains. During the prediction phase, a spectral gaussian grid compensation structure is further integrated to enhance the accuracy of predictions for the target bands. Ultimately, by compressing the residual information based on the differences between the predicted bands and the true bands, an efficient representation of the residual information in the compressed domain is achieved. The predicted image is then fused with the compressed residual to obtain a more expressive and potentially compressed representation. During the decoding phase, the diffusion process of the conditional diffusion model (CDM) utilizes compressed representations as potential conditions to guide the model in progressively reconstructing images over multiple time steps. Experimental findings demonstrate that the bi-residual compression network achieves superior peak signal-to-noise ratio (PSNR) across nine datasets, with an improvement of approximately 1 dB over SOTA methods. Moreover, BRC-CDM attains a PSNR that surpasses that of the majority of current methodologies, while also delivering enhanced spectral fidelity and perceptual quality. To foster reproducibility and further development, we release the full implementation of BRC-CDM at https://github.com/Nicle-L/BRC-CDM.

Hyperspectral images play an important role in the field of remote sensing, and similar to ordinary RGB-based images, reconstructing the original information with higher quality using fewer bits is an essential task. Most existing hyperspectral image compression methods utilize a transform-based compression framework that reconstruct the original image after converting the input into a latent representation with a specific size. This kind of approaches have achieved some success, however, they suffer from two problems. First, the encoder and decoder used for transformation take up a huge amount of computational resources, both for training and deployment. Second, the upper performance limit is not satisfactory, that is to say, huge computational cost does not bring a matching performance gain. Based on this, we propose a novel hyperspectral image compression method. Specifically, we employ neural radiance fields (NeRF) to compress hyperspectral images, and unlike transform-based methods, the proposed method encodes the hyperspectral coordinate information, which is fitted to hyperspectral pixel values using multilayer perceptrons (MLPs). After that, we only need to compress the weights of the generated MLPs using the model compression method (in this paper, we only employ the weight quantization) to efficiently save the hyperspectral images, since the MLPs model, working as the fitting function, can be regarded as the compressed representation of the hyperspectral image. At the same condition, the proposed method achieve nearly 5 dB higher in PSNR and 6 dB higher in MS-SSIM than the comparison deep-learning based methods.

In recent years, the impressive feature representation capabilities of deep learning have opened up new possibilities for image compression. Most of the existing learning-based image compression techniques rely on convolutional neural networks (CNNs) to obtain local feature representations of the input image using moving windows. However, the convolutional kernel of CNNs only considers local spatial relationships in the perceptual field, while ignoring long dependencies in the features. This results in incomplete compression of the latent representation. The similar features in remote sensing (RS) images are more abundant and widely available, and thus, the inconvenience of CNNs is more obvious. To address this problem, we propose a hybrid attention compression network (HACN) for RS images, which can exploit long dependencies in the latent representation to achieve a more compact bitstream. Specifically, the residual attention module (RAM) and graph attention module (GAM) are attached to the network. This hybrid attention mechanism (HAM) helps the encoder extract spatial and cross-channel long dependencies in the feature transformation process, which in turn improves the rate distortion metric. Meanwhile, regarding the computational cost of the network, we also propose a light GAM, which greatly reduces the burden of compression. The experimental results demonstrate that the proposed method achieves satisfactory rate-distortion performance, compared to conventional and CNN-based methods.

Benefiting from the excellent texture recovery capability of generative adversarial networks (GANs), generated images are capable of maintaining clear texture features even when compressed into extremely low-bit streams. In recent years, the GAN has made great progress in extremely low-bit compression for natural images. However, a few studies have been conducted on extremely low-bit compression for the remote-sensing (RS) field. We find that a single GAN tends to generate visually pleasing texture information, and this characteristic may affect the visual effect and accuracy of other computer vision tasks. Therefore, we propose a coupled compression generation network (CCGN) that reconstructs the image content and detailed textures separately and fuses them to achieve a balanced image reconstruction task at extremely low bitrates for remote-sensing images. Specifically, a multidimensional residual attention mechanism (MRAM) is adopted to achieve extremely low-bit stream generation, whereas contentwise images and texturewise images are reconstructed using the same generator with different training strategies. We further optimize the texture generation strategy, and an enhanced perceptual-guided refinement stage (EPGRS) and a multiscale fusion discriminator (MSFD) are developed for a more realistic texture. The proposed method achieves outstanding results on compression tasks on the dataset for object detection in aerial images (DOTA), and the fused results of extremely low-bit streams also perform well in object detection tasks, significantly alleviating pressure from bandwidth and storage space.

The accurate imaging of stationary human targets in the indoor scene containing strong scatterers such as cabinets, tables, and chairs is very important for the through-the-wall radar (TWR) system. The convolution neural network (CNN) has been used to enhance radar imaging quality. In this letter, a novel multiresolution fusion network (MRFN) based on the deep layer aggregation (DLA) method is proposed for TWR imaging. The proposed MRFN can accurately localize the weak scattering human targets and provide the scattering intensity differences between the strong and weak targets. Both the simulated and real TWR data are used to evaluate the imaging performance of the proposed MRFN. The experimental results demonstrate the superiority of the proposed TWR imaging method over the existing CNN-based imaging methods.

Deep learning algorithms have been widely used in ship detection with synthetic aperture radar (SAR). However, the complex background, clutter noise, and large span of ship sizes have adverse effects on the feature extraction, which seriously limits the ship detection accuracy. To address this issue, a cross-scale regional prediction-aware network (CSRP-Net) is developed to advance the ship detection performance in SAR images. First, the cross-scale self-attention (CSSA) module is designed to suppress the influence of noise and complex backgrounds and enhance the ability to detect multiscale targets. Furthermore, a regional prediction-aware one-to-one (RPOTO) label assignment is proposed to select the foreground samples more conducive to classification and regression in the training stage. Extensive experiments have proved that the designed method can significantly improve the detection performance against several start-of-the-art algorithms on two classical benchmark datasets.

Synthetic aperture radar (SAR) image compression plays an important role in the manipulation of images. However, existing optical compression methods cannot properly handle SAR compression due to the absence of feature learning and representation for SAR images. In this letter, we propose an end-to-end trainable model to effectively fit the feature distribution and reduce information dependencies toward SAR image compression. To better parameterize the distribution of latent codes, a discretized Gaussian adaptive model is designed to achieve a flexible entropy process. To further remove the remaining redundancies, generalized subtractive normalization is introduced to reduce the statistical dependencies in SAR images. Extensive experiments show that the proposed compression method outperforms the traditional compression methods and learning-based algorithms on both the ICEYE and Sandia datasets.

Despite deep convolutional neural networks (DCNNs) having been used extensively in radar-based human activity recognition in recent years, their performance could not be fully implemented because of the lack of radar dataset. However, radar data acquisition is difficult to achieve due to the high cost of its measurement. Generative adversarial networks (GANs) can be utilized to generate a large number of similar micro-Doppler signatures with which to increase the training data set. For the training of DCNNs, the quality and diversity of data set generated by GANs is particularly important. In this paper, we propose using a more stable and effective Wasserstein generative adversarial network with gradient penalty (WGAN-GP) to augment the training data set. The classification results from the experimental data have shown the proposed method can improve the classification accuracy of human activity.

Estimating unknown wall parameters is of great importance for the application of through-the-wall radar (TWR). The time-delay-only estimation (TDOE) method is able to efficiently retrieve the constitute parameters of the homogeneous wall under test. For the TDOE method, the estimation accuracy of time delays associated with wall reflections directly affects the accuracy of wall parameters estimation. In this paper, we propose a sparse recovery method for estimation of wall parameters, which utilizes the orthogonal matching pursuit (OMP) algorithm to perform the time delay estimation of echoes backscattered from the wall at each antenna separation. Numerical simulation results have shown that the proposed estimation method is capable of providing the estimation of wall parameters with higher accuracy.

High-resolution diffraction tomographic (DT) ground-penetrating radar (GPR) image formation requires the use of wideband signal and large antenna array aperture, which leads to the generation of large amounts of imaging data. A compressive sensing multibistatic GPR DT imaging algorithm is presented in this letter. The proposed imaging algorithm can provide the advantage in terms of reducing the measured data in the frequency domain while maintaining the image quality of the reconstructed scenario. The imaging results reconstructed via the processing of synthetic data have verified the validity and effectiveness of the proposed imaging method.

Estimating the time delays of buried target echoes is particularly important for the application of ground penetrating radar (GPR). Due to its smaller computational burden, the estimation of signal parameters via rotational invariance technique (ESPRIT) is preferred to process the buried target echoes in the frequency domain and to obtain the accurate super-resolution time delays. In this letter, we give an in-depth analysis of the essential preprocessing steps for the application of ESPRIT to practical GPR measurement data. In particular, an improved spatial smoothing method is adopted to construct the correlation matrix for the robustness of the time-delay estimation result. The effectiveness of the algorithm is verified by the synthetic data from a horizontally stratified medium model using the finite-difference time-domain method, which explicitly takes into account surface scattering for a more realistic scenario.