The target detection task of ground penetrating radar (GPR) based on deep learning has received widespread attention. Previous studies focused more on the features of targets in images and achieved excellent performance. However, in the practical application of these methods, GPR image is cut into several slices for feeding to the model for training and inference, which not only requires accumulating pulses but disrupts the continuity of pulse information, making it difficult to communicate semantic information of different parts of the same pulse in different slices. To address these issues, this letter proposes a pulse-by-pulse target detection model, namely, 1-D mix attention (MA) TransUnet, for GPR, avoiding pulse accumulation and preserving the continuity of pulse information. In structure, the spatial and channel mixed attention mechanism replaces skip connections in 1-D Unet, which effectively enhances the target features in pulse data. In addition, transformer block (TB) based on multihead self-attention (MSA) is applied to the downsampling feature map of 1-D Unet, which allows the model to effectively understand the global semantic information and suppress nontarget features that are similar to the target features in pulse data. Finally, the effectiveness of 1-D MA TransUnet is validated using GPR pulse data containing steel mesh as a case study. The model achieved an accuracy of 83.07%, a recall rate of 71.64%, and an F1-score of 76.93%, respectively.

GPR clutter removal significantly benefits subsequent target recognition, detection, and imaging, enhancing the subsequent processing quality. Traditional clutter removal approaches can only remove noise in simple environments. To solve this problem, We proposed a novel improved triplet attention u-net(ITU-Net) which focuses on the hyperbolic feature w e need while disregarding irrelevant ground clutter and other background noise. The ITU-Net network enhances image reconstruction capability and facilitates rapid image processing. The improved triplet attention module captures cross-domain interaction between any two domains between H, W, and C and considers long-distance dependencies separately in H, W, and C. The experimental results demonstrate that we can effectively retain the information of the hyperbolas while eliminating noise in complex environments.

Scattering imaging is a pervasive scenario in many areas, especially challenging the performance of remote sensing and automatic target recognition (ATR). Recently, deep learning was utilized for synthetic aperture radar (SAR) ATR in scattering scenarios by extracting the feature of speckle patterns. However, huge computational costs and power consumption challenge its development. Here, we develop a speckle-based residual optronic convolutional neural network (S-ROPCNN) for SAR target recognition. Specifically, we model the light scattering scenarios and build the optical imaging system to produce the speckle patterns for network training. The S-ROPCNN performs SAR target recognition in optical platforms with the speed of light, low computational cost, and low energy consumption. Experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset demonstrate the feasibility of S-ROPCNN for SAR target recognition in scattering imaging scenarios.

Marine surveillance radar can monitor the maritime environment under all weather conditions, but the presence of sea clutter significantly impacts its target detection performance. To effectively mitigate the influence of sea clutter on radar imaging, this paper proposes a neural network based on DenseNet combined with wavelet transform. The wavelet transform, known for its reversibility that preserves the original image information, is capable of extracting features at different levels of detail. DenseNet enhances the flow of extracted features, effectively alleviating the issues of gradient explosion or vanishing. The network is evaluated using data collected by the IPIX radar as the sea clutter noise dataset. Under various input conditions with different clutter-to-signal ratios, the proposed network achieves an average improvement of 18.78dB in clutter-to-signal ratio. Experimental results demonstrate the network’s effectiveness in suppressing sea clutter noise.

For reducing computational burden of electronic hardware and increasing practical recognition performance of optron-ic convolutional network (OPCNN), here we propose an optical backpropagation algorithm and realize the in-situ training OPCNN in optical platform for SAR target recognition. Training networks according proposed algorithm, major computational operations in forward and backward propagating process are all executed in optics with the speed of light and low consumption. Several experiments on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset demonstrate the feasibility of proposed in-situ training algorithm.

Due to the atmospheric turbulence, defocusing, noise and other factors, the optical remote sensing image acquisition may become blurred. Therefore, it is critical of deblurring the images by algorithm. In recent years, neural network algorithms have shown excellent performance in optical re-mote sensing images deblurring. However, neural network algorithms have some limitations at the same time. They lack interpretability and need large amounts of training samples. The traditional deblurring algorithms are interpretable, but the performance is not as good as the neural network algorithms. In order to obtain an interpretable deblurring algorithm with good performance, this paper proposes a deblurring algorithm based on deep unfolding method, which is the combination of traditional algorithms and neural networks. It can achieve good performance and be interpretable at the same time. We demonstrate the effectiveness of the algorithm on remote sensing datasets with PSNR values and visual deblurring images. The experiments show the proposed algorithm has better deblurring results.

This paper presents a multi-structure extraction kernel dictionary learning (MSEK-DL) method for synthetic aperture radar (SAR) automatic target recognition (ATR). In order to extract the multi-structure features of SAR images for data enhancement and noise suppression, a matrix approximation method is used. Instead of using traditional linear dictionary learning method, non-linear kernel function is used to map the targets into a high-dimensional space, in order to obtain a better classification performance. The training method and optimization steps of MSEK-DL are presented in this paper. We carried out the experiment based on MSTAR dataset to demonstrate the effectiveness of the proposed classification algorithm. The experimental results show that the classifi-cation algorithm has better classification performance than some representative dictionary learning algorithms, espe-cially for small training datasets.

Single image super-resolution technology is critical in remote sensing fields because it can effectively improve the details of target images. However, the application of deep learning is limited due to the lack of interpretability and the need for many parameters. This letter proposes an interpretable dual-branch multi-scale channel fusion unfolding network (DMUNet) for optical remote sensing image (ORSI) super-resolution. We design an unfolding network with double branches, each optimized with different strategies. Two branches focus on texture and edge reconstruction, respectively. This unfolding network follows the iteration process of the alternating direction method of multipliers (ADMM) and can learn the hyper-parameters adaptively. The functions of the two branches can complement each other. Further, to better fuse the feature maps of the two branches, a multi-scale fusion module is proposed. This module can effectively fuse information between different branches, scales, and channels. It is noted that it only requires a little computation cost. Experiments on two public ORSI datasets demonstrate that our method can achieve significant performance in both quantitative evaluation and visual results.

Single image super-resolution technology is critical in remote sensing, effectively improving the resolution of target images, with super-resolution algorithms based on deep learning demonstrating superior performance. However, most neural networks present shortcomings, such as lack of interpretability and requiring a long training time, limiting them in some application scenarios. Moreover, due to multi-degradation factors, tasks put forward higher requirements for the adaptability of algorithms. Therefore, this work develops a structured deep unfolding network (SDUNet), which is adaptable and requires a lower training time by cascading multiple small network modules. Additionally, the unfolding strategy proposed deals with multiple degradations, fully exploiting prior knowledge. The suggested method is challenged against state-of-the-art neural network methods on one optical remote sensing image dataset and one natural image dataset. The experimental results demonstrate our method’s effectiveness in requiring less training time, involving fewer parameters, and achieving a higher reconstruction performance for optical remote sensing image super-resolution.

Single image super-resolution technology based on deep learning is widely applied in remote sensing. In recent years, the deep unfolding super-resolution strategy has been proposed, which combines the neural networks with traditional optimization-based algorithms, making the neural networks interpretable and achieving high performance. However, the typical deep unfolding algorithms usually treat different kinds of blurring kernels in the same way, so the algorithms cannot take advantage of the properties of blurring kernels, limiting the algorithm’s performance. To design a super-resolution network that can fully use the properties of Gaussian blurring kernels, a dual-resolution local attention unfolding network (DLANet) is proposed. Based on the Gaussian blurring functions, a low-resolution (LR) space branch is designed to supplement the high-resolution (HR) space branch. Specifically, for Gaussian blurring kernels, the closer the pixel is to the center, the greater the weight is. It means that the pixel points retained after downsampling will contain more information about the original corresponding pixel points, and it could be easier to estimate their original pixel values. So we design two branches. The HR branch completes the estimation of the whole image, and the LR branch only estimates the points retained after downsampling. To better complete the feature fusion of the two branches, we propose a row-column decoupling local attention module. This module can retain more information when fuse features and the row-column decoupling strategy can reduce computational complexity. Comprehensive experiments demonstrate the superiority of our method over the current state-of-the-art on remote sensing datasets.