Deep learning-based detection and tracking of fin whales using high-resolution space-borne remote sensing data

Despite the ban on commercial whaling, the conservation of fin whale populations remains a significant challenge due to human-induced threats such as ship collisions, fishing gear entanglements, and underwater noise pollution. Traditional monitoring methods are logistically challenging and expensive, especially in remote and inaccessible regions. Recent advancements in high-resolution satellite imagery have demonstrated potential for automated marine species monitoring; several research gaps remain, including limited spectral band utilization, suboptimal deep-learning model adaptations, and lack of real-time tracking capabilities. This study presents an advanced deep-learning framework integrating U-Net for semantic segmentation, an enhanced YOLO model for object detection, and ResNet101 for classification to automate the detection and tracking of fin whales in satellite and infrared imagery. A key contribution is the integration of specific spectral bands optimized for underwater visibility, improving detection accuracy. The proposed system is deployed on edge devices, enabling real-time fin whale tracking with geospatial mapping of their locations. Experimental results demonstrate high performance across multiple datasets. U-Net achieves a segmentation accuracy of 92.21 %, the enhanced YOLO model attains a mean average precision (mAP) of 82 %, and ResNet101 reaches a classification accuracy of 99 %. Comparative analysis against existing methodologies highlights the improved detection precision and robustness of the proposed approach. By addressing key research gaps in spectral band selection, deep learning adaptation, and real-time deployment, this work contributes significantly to automated marine species monitoring and conservation. This study integrates drone-based surveys, hyperspectral imaging, thermal imagery, and Google Earth data with satellite imagery to enhance tracking capabilities.