Real-time remote sensing segmentation technology is crucial for unmanned aerial vehicles (UAVs) in battlefield surveillance, land characterization observation, earthquake disaster assessment, etc., and can significantly enhance the application value of UAVs in military and civilian fields. To realize this potential, it is essential to develop real-time semantic segmentation methods that can be applied to resource-limited platforms, such as edge devices. The majority of mainstream real-time semantic segmentation methods rely on convolutional neural networks (CNNs) and transformers. However, CNNs cannot effectively capture long-range dependencies, while transformers have high computational complexity. This paper proposes a novel remote sensing Mamba architecture for real-time segmentation tasks in remote sensing, named RTMamba. Specifically, the backbone utilizes a Visual State-Space (VSS) block to extract deep features and maintains linear computational complexity, thereby capturing long-range contextual information. Additionally, a novel Inverted Triangle Pyramid Pooling (ITP) module is incorporated into the decoder. The ITP module can effectively filter redundant feature information and enhance the perception of objects and their boundaries in remote sensing images. Extensive experiments were conducted on three challenging aerial remote sensing segmentation benchmarks, including Vaihingen, Potsdam, and LoveDA. The results show that RTMamba achieves competitive performance advantages in terms of segmentation accuracy and inference speed compared to state-of-the-art CNN and transformer methods. To further validate the deployment potential of the model on embedded devices with limited resources, such as UAVs, we conducted tests on the Jetson AGX Orin edge device. The experimental results demonstrate that RTMamba achieves impressive real-time segmentation performance.

With the increasing demand for automatic underwater object detection, different deep learning (DL) methods have been performed on forward-looking sonar (FLS) images in recent years. However, DL-based approaches face challenges of overfitting due to excessive parameters and limited generalization ability. This study aims to develop a lightweight FLS detection network (FLSD-Net) that improves the detection of small underwater targets by addressing overfitting and generalization issues. The FLSD-Net incorporates three key modules - a light initial downsampling module (LID) to reduce information loss, a lightweight feature extraction module (LFE) to reduce overfitting from parameter redundancy, and an optimized network structure prioritizing high-resolution shallow layers to improve detection accuracy and speed. Validation on public and self-created datasets shows that the proposed method achieved a balance of detection speed and accuracy with better performance than state-of-the-art (SOTA) lightweight networks. In addition, the results in generalization experiments exhibit reduced missed and false detections, indicating improved robustness. This research advances underwater target detection by developing a lightweight network with improved generalization, providing a robust solution for real-world applications.

Moving target detection in the disturbing environment has been significantly active and challenging for underwater acoustic array signal processing. One of the most effective approaches is based on low-rank matrix estimation, which divides the reconstructed imaging frames into low-rank and sparse matrix components, with the low-rank structure representing the interference and the sparse component corresponding to the moving target. However, the array model has not been well explored, taking the reconstruction and moving target separation as independent steps. In this paper, the sparse Bayesian learning (SBL) approach is employed to recover the stable interference and the moving target simultaneously and separately, retaining the high reconstruction quality pertaining to SBL. Specifically, in each frame, a shared prior distribution is assigned to the stable interference, while the time-varying component (the moving target) is modeled with a unique prior distribution. The hierarchical model is solved by the variational Bayesian inference. The superior performance of the proposed algorithm is validated by simulation and experimental data in an underwater acoustic test tank and Songhua Lake in China, compared with state-of-the-art low-rank matrix estimation algorithms.

Direction of arrival (DOA) and source amplitude estimation are widely encountered in detection and communication as traditional array processing problems. Recently, sparse Bayesian learning (SBL) technique is increasingly exploited, which outperforms conventional compressive sensing (CS) approaches and provides an improved alternative for DOA and amplitude estimation. Although the accuracy of DOA is enhanced somewhat through SBL, the on-grid implementation of SBL suffers the grid mismatch errors due to predefined grids. The off-grid SBL can be employed to address the issue. However, the computation load of off-grid SBL is heavy because of per-iteration matrix inversion and slow convergence, which restricts its application for online systems. An off-grid space alternating multi-snapshot SBL (OGSA-MSBL) approach used for DOA and source amplitude estimation is proposed in this paper, which introduces the offset between the ground-truth DOA and the grid nodes to refine the model, and employs the space alternating (SA) approach to solve the problem efficiently. Compared with the original SA, the DOA accuracy can be improved even with a reduced number of grids, while the source amplitude can be estimated with a higher precision with the first-order Taylor refined beamforming model. As shown by the simulation and experimental results, the recovery accuracy of the proposed method is superior to the state-of-the-art on-grid SBL approaches and comparable to its off-gird counterpart, while much faster in terms of computation efficiency.