Oriented ship detection in synthetic aperture radar (SAR) images is essential for maritime surveillance, but remains challenging because obtaining reliable rotated bounding box (RBox) annotations is time-consuming. In this paper, we propose a collaborative prior and confidence learning network (PCLNet) for point-supervised oriented ship detection. Single-point annotations are cost-effective but provide limited geometric cues, leading to inaccurate localization and unstable pseudo-label quality. PCLNet addresses this limitation with a teacherstudent framework that stabilizes pseudo-label learning and improves target-domain reliability. The network further incorporates three modules: reliability-aware confidence learning (RACL), saliency-guided label refinement (SLRM), and a two-stage pseudo rotated bounding box refinement (PRR) process. The RACL predicts an intersection over union quality score for each proposal and integrates it with class confidence to align classification and localization. The SLRM aggregates multilevel features to generate a SAR-aware saliency prior that highlights ships and suppresses nontarget interference. This prior provides spatial guidance for localization and pseudo-label refinement during teacher–student training. Guided by this prior, PRR progressively refines coarse pseudo RBoxes into high-quality pseudo labels and improves point–RBox consistency. Experiments on two SAR ship datasets demonstrate competitive accuracy and strong robustness in both inshore and offshore scenes while requiring only single-point annotations.
The spread of the invasive species Spartina alterniflora poses a serious threat to coastal salt marsh ecosystems, requiring accurate, scalable monitoring solutions. We introduce the Normalized Difference Spartina Phenology Index (NDSPI), a phenology-based metric derived from the Normalized Difference Phenology Index, to improve discrimination of S. alterniflora using Sentinel-2 time-series imagery. Phenological curves for dominant salt marsh vegetation-S. alterniflora, Phragmites australis, Suaeda salsa, and mixed weeds-were reconstructed via Harmonic Analysis of Time Series. The Jeffries-Matusita distance was employed to identify key phenological stages and spectral indices, guiding the development of NDSPI to exploit species-specific temporal patterns. Using NDSPI combined with K-means++, the spatial distribution of multiple salt marsh vegetation types in the Linhong Estuary of Haizhou Bay from 2019 to 2023 was mapped, with S. alterniflora achieving a precision of 0.82, a recall of 0.88, and an F1-score of 0.85. Comparative experiments confirmed the innovation of NDSPI, demonstrating its superior capability to amplify target separability over traditional NDVI-based and supervised machine learning methods. Validation experiments in the Dongtan wetland of Chongming Island demonstrated strong generalization. By overcoming the reliance on extensive training data and mitigating background noise, NDSPI offers an effective, highly automated tool for large-scale monitoring of S. alterniflora invasions in complex coastal wetlands.
Remote sensing-assisted multimodal large language models provide a powerful approach for spatial semantic perception in applications such as intelligent transportation and urban mobility analysis. Although multimodal large language models (MLLMs) have recently advanced remote sensing scene understanding, existing models mainly focus on general remote sensing semantics and lack specialized designs for trajectory-related traffic semantic understanding and traffic-scene reasoning. To bridge this gap, we propose MM-RSTraj, a remote sensing-assisted multimodal framework tailored for trajectory-traffic semantic understanding while simultaneously enhancing general spatial semantic perception. Built upon the LLaVA-OneVision architecture, MM-RSTraj adopts a two-stage multimodal instruction tuning strategy to progressively enhance cross-modal semantic interaction and traffic-scene semantic reasoning from remote sensing imagery. To support this process, we construct two high-quality instruction datasets: RSI-Instruct, which extends RSICap with multiturn instruction-response supervision for general remote sensing semantic understanding; and RSI-Traffic, specifically designed for trajectory-related traffic semantic understanding, focusing on road topology, traffic-scene semantics, urban spatial structures, and functional-region relationships. Experimental results demonstrate that MM-RSTraj achieves strong performance in both general remote sensing semantic tasks, including remote sensing image captioning and remote sensing visual question answering, and trajectory-traffic semantic understanding tasks. In addition, MM-RSTraj exhibits stronger capability in traffic-scene reasoning, semantic organization, and high-level traffic semantic understanding.
Remote sensing image dehazing is an important low-level vision task for improving image interpretability and supporting downstream applications such as land-cover recognition, object detection, and change detection. Existing CNN-based and Transformer-based methods have achieved remarkable restoration performance, but their computational cost remains a concern in practical remote sensing applications. To address this issue, this paper proposes a hierarchical spiking restoration network, termed HSRNet, for remote sensing image dehazing. HSRNet introduces spiking neural networks into the dehazing framework and exploits temporal spiking representations to model haze-degraded structures. Specifically, a hazy remote sensing image is first replicated into a multistep temporal sequence, allowing leaky integrate-and-fire neurons to accumulate degradation-sensitive responses over time. A hierarchical multiscale spiking network is then employed to extract and restore structural information across different spatial resolutions. To enhance feature representation, a spiking feature enhancement block is introduced, which combines spiking transformation, residual feature propagation, and adaptive response refinement to strengthen haze-weakened details. Finally, a temporal refinement reconstruction head aggregates temporal responses and reconstructs the haze-free image in the continuous domain. Extensive experiments on benchmark datasets demonstrate that HSRNet achieves favorable quantitative and qualitative performance while maintaining efficient model complexity. These results indicate that hierarchical spiking representation provides a promising direction for remote sensing image dehazing.
Dense terrestrial LiDAR surveys produce high-point-count point clouds that challenge processor throughput and memory constraints of online perception pipelines. Standard tree-based indices often suffer from poor cache locality due to pointer-heavy traversals. To address this, we present C-Index, a LiDAR-aware, key-based indexing backend designed for cache-friendly exact radius search. C-Index exploits the sensor's native geometry by decomposing 3D space into a beam-aligned vertical constraint and a compact horizontal complex-form representation (C=rho sin theta e(i phi)). These coordinates are linearized into a 64-bit space-filling curve key, enabling contiguous memory layout and efficient range scanning without complex tree structures. For radius queries, the method guarantees exactness via a two-stage process: conservative candidate generation followed by Euclidean-distance verification, whereas kNN retrieval is supported only as a practical retrieval mode and is not included in the exactness claims of this paper. Evaluated on the UrbanGreen-Mix dataset, C-Index reduces radius-query latency to 25.8 mu s-yielding substantial latency reductions over the compared in-memory baselines-while maintaining a competitive memory footprint of 210.3 MB. Furthermore, in a downstream vegetation classification pipeline, replacing the indexing backend with C-Index reduces end-to-end processing time by 80% (180.5 to 35.2 s) without compromising accuracy. C-Index thus provides a high-performance, plug-and-play primitive for LiDAR processing workflows requiring exact geometry.
