This study introduces a hierarchical dual-model detection framework for accurately monitoring spotted seals (Phoca largha) in the Liaohe River estuary by using deep learning on Unmanned Aerial Vehicles (UAVs). To address challenges such as weak target features, background interference and limited edge computing capacity, this study deploys an optimized FF-YOLOv10 lightweight model on UAVs for rapid target localization, followed by an enhanced PP-YOLOv7 model on ground stations for precise detection. The FF-YOLOv10 model reduces computational complexity by 24.2% and increases inference speed by 33.3%, while the PP-YOLOv7 model achieves 94.2% precision with a 1.9% increase in recall rate. This framework provides an efficient and precise technical solution for the long-term ecological monitoring of marine endangered species, supporting habitat conservation policy formulation and ecosystem health assessments.

To address the need for cost-effective imagery in precise remote sensing monitoring of coastal aquaculture ponds, a super-resolution method based on a dynamic sparse self-attention mechanism is proposed. The method enhances feature extraction by selectively focusing on the most relevant areas of the image, while maintaining low computational cost. Built on the Transformer architecture, it incorporates interval-based Top-K feature selection and a scale-aware interaction module to better capture multi-scale pond features. Experimental results demonstrate that the proposed approach outperforms existing methods, achieving superior performance in both Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). Specifically, PSNR values of 27.25 dB and 29.42 dB are achieved for simulated datasets from Qingdao and Yantai, respectively, while real Sentinel-2 imagery benefits from enhanced texture preservation and noise reduction. This approach provides a promising solution for cost-effective, high-precision monitoring of coastal aquaculture ponds, contributing to the sustainable development of aquaculture resources. The codes are available at https://github.com/Lrange-212/CAPSR.git.

The existing pre-defined coverage path strategies in unmanned aerial vehicle (UAV)-based vertical photogrammetry are inadequate for the monitoring and surveying of spotted seal populations with dynamic group distribution characteristics. This paper presents a novel real-time, perception-driven adaptive coverage path planning (CPP) approach utilizing UAV edge computing for intelligent recognition. The proposed method integrates onboard target detection, multi-object tracking (MOT), and visual positioning during UAV patrol operations to achieve real-time perception of individual locations within mobile populations. To mitigate false detections and re-detections, this study proposes a Spatial Noise Filtering (SNF) method for UAV-based detection points. A three-level noise removal scheme is designed, consisting of a distance threshold, low-density point filtering within core regions, and global low-density filtering, enabling systematic suppression of spatial noise generated during detection. On this basis, the SNF method is integrated with a Genetic Algorithm (GA) to achieve dynamic optimization of the UAV coverage path. Experimental results show that the adopted YOLOv11 model achieves a precision of 0.9. The multi-object tracking (MOT) technique based on UAV video streams reduces redundant positioning by 61% (from 389 detection points to 150 tracking points), and the average visual positioning error is 4.9 m at flight altitudes below 75 m. The proposed SNF method effectively removes most invalid coverage regions, significantly improving the spatial reliability of the detection results. Furthermore, the GA-based optimization of the UAV coverage path enables adaptation to the uncertain spatial distribution of spotted seal populations, meeting the requirements for dynamic population surveying and real-time monitoring.