Deep learning has become increasingly popular in hyperspectral image (HSI) and light detection and ranging (LiDAR) data classification, thanks to its powerful feature learning and representation capabilities. However, HSI often contains substantial redundant information, which can hinder efficient data utilization. Furthermore, the significant disparity in information content between HSI and LiDAR data poses a major challenge in representing and aligning semantic information across these two modalities. To address these challenges, we propose a fusion network structure guided by feature reconstruction embedding. This approach employs feature decomposition to reconstruct HSI features and incorporates weight embedding to seamlessly integrate the reconstructed information into classification features. Furthermore, we introduce a cross-modal attention fusion module designed to merge extracted HSI and LiDAR features. This module fully exploits the complementary nature of these two type of feature, facilitating effective information exchange and semantic alignment across multimodal data. We evaluated our method on three widely used HSI and LiDAR datasets: Houston 2013, Augsburg and MUUFL. Experimental results demonstrate that our proposed FRGFNet significantly outperforms traditional probabilistic methods and state-of-the-art deep learning networks, showcasing its effectiveness in multi-source data fusion.

Onboard processing for remote sensing satellites has long been of interest in ocean monitoring because it can provide timely information on rapidly changing ocean environments. Recently, the integration of artificial intelligence (AI) into onboard processing has emerged as an active research topic. However, application scenarios and on-orbit demonstrations of AI-based onboard processing remain limited. In this study, we achieved the first on-orbit detection of ocean eddies using AI on the Haishao-1 (HS-1) spaceborne synthetic aperture radar (SAR) satellite, marking the first attempt to monitor an ocean dynamic process directly onboard. Specifically, an onboard application module was developed for HS-1, integrating image preprocessing and deep learning-based detection based on our previously developed Eddy Detection based on YOLO (EOLO) model. On-orbit experiments verified stable execution of the application module. Ground reproduction of the tasks using hardware identical to the onboard computing units yielded an average processing time of 3.4-6.1 s per image. Evaluation against expert annotations of 288 HS-1 scenes showed that EOLO, originally developed using Sentinel-1 imagery, achieved an F1 score of 78% after threshold adjustment. The application module has also been deployed on the Haishao-2 platform for Arctic sea ice extraction, further demonstrating its extensibility. This study presents the implementation and on-orbit validation of the onboard application module, providing a practical reference for future applications.

Ocean submesoscale eddies, characterized by horizontal scales less than the first baroclinic Rossby radius of deformation, are increasingly recognized for their critical roles in marine ecosystems, ocean energy balance, and the Earth’s climate system. Despite extensive research on these submesoscale features through high-resolution simulations and regional observations, our knowledge of their dominant driving mechanism from a global perspective is very limited. Here, we present the global observational evidence for the primary role of mesoscale Lagrangian coherent structures (LCSs) in driving submesoscale eddy generation by synergic application of high-resolution spaceborne synthetic aperture radar data and radar altimeter data. Applying a deep-learning detection system to more than three million global Sentinel-1 and Envisat SAR images, we found that more than 80% of detected submesoscale eddies are clustered within a 10-km range around mesoscale LCSs characterized by high kinetic energy and persistent straining. Further composition analysis quantifies that more than half of submesoscale eddies occur within the ring-shaped areas of coherent mesoscale eddies, where the strong strain is conducive to frontogenesis. Our findings highlight that the generation of submesoscale eddies is attributed to instabilities initiated by strain-induced frontogenesis. This study establishes a new paradigm for locating submesoscales by targeting LCSs, thereby supporting a global evaluation of their contributions to energy balance and material transport.

Eddies are ubiquitous in the Arctic Ocean, as observed by spaceborne synthetic aperture radar (SAR) over the open seas and marginal ice zones (MIZs). In MIZs, SAR captures eddies mainly owing to the accumulation of drifting ice floes, known as ice eddies. We recently proposed a deep learning network named EOLO (Eddy detection based on You-Only-Look-Once algorithm) to detect ocean eddies by SAR, which presents high precision (about 95%) on open-water eddies in generalized experiments; however, it will over-extract complex background information of sea ice and generate false detections. Therefore, we move further for ice eddy detection in this study. First, a Sentinel-1 (S1) SAR data-based ice eddy dataset was established, placing particular emphasis on incorporating diverse background (non-eddy) samples, especially for sea ice; subsequently, the algorithm for ice eddy detection was trained based on the novel YOLOv8 network, named EOLO-ice. Finally, incorporating full and slicing-aided inferences, the EOLO and EOLO-ice were integrated to identify eddies in both open oceans and MIZs. Extensive testing was performed on EOLO-ice using the ice eddy dataset, with a precision (P) of 92.37% and a recall (R) of 90.23%. The effectiveness of EOLO-ice for ice eddy detection was further illustrated by visualizing feature maps at critical stages, highlighting the crucial role of the C2F module and decoupled head in YOLOv8 to suppress background features of sea ice in SAR imagery. Finally, the integrated eddy inference framework was applied to 912 S1 scenes acquired over the Greenland Sea and Fram Strait, and then achieved a high performance, with a P and a R of 94.39% and 89.46%, respectively, suggesting its good applicability in the open seas and MIZs.

Ocean internal waves and eddies are typical mesoscale and submesoscale processes that play critical roles in ocean energy transfer and material transport. Their complex morphology and varying background conditions make detection from Synthetic Aperture Radar (SAR) imagery challenging. In this paper, we propose a novel multi-scale YOLO framework, termed MuLO, which designed to investigate the intrinsic connections among multi-scale ocean dynamic processes and to enable collaborative detection of oceanic eddies and internal waves in SAR imagery. This model integrates a Dynamic Grouping Cascade Module (DGCM) and a Multi-Scale Contextual Aggregation (MSCA) mechanism to enhance weak feature extraction and multi-scale feature fusion. A multi-source spaceborne SAR dataset including X-, C-, and L-band imagery acquired in the South China Sea was constructed, and a multi scale slicing inference strategy was designed for practical analysis. Experiments show that MuLO, with only 2.41M parameters, achieves 90.8% mAP50, outperforming several state-of-the-art models. Multi-source evaluation further demonstrates robust detection across radar bands, with a mean average precision (mAP) of 87.8% in simultaneous eddy and internal wave identification. These results highlight the potential of MuLO for operational monitoring of ocean dynamic processes by spaceborne SAR.