Enhanced Fishing Monitoring in the Central-Eastern North Pacific Using Deep Learning with Nightly Remote Sensing

The timely and accurate monitoring of high-seas fisheries is essential for effective management. However, efforts to monitor industry fishing vessels in the central-eastern North Pacific have been hampered by frequent cloud cover and solar illumination interference. In this study, enhanced fishing extraction algorithms based on computer vision were developed and tested. The results showed that YOLO-based computer vision models effectively detected dense small fishing targets, with original YOLOv8 achieving a precision (P) of 89% and a recall (R) of 79%, while refined versions improved these metrics to 93% and 99%, respectively. Compared with traditional threshold methods, the YOLO-based enhanced models showed significantly higher accuracy. While the threshold method could identify similar trend changes, it lacked precision in detecting individual targets, especially in blurry scenarios. Using our trained computer vision model, we established a dataset of dynamic changes in fishing vessels over the past decade. This research provides an accurate and reproducible process for precise monitoring of lit fisheries in the North Pacific, leveraging the operational and near-real-time capabilities of Google Earth Engine and computer vision. The approach can also be applied to dynamic monitoring of industrial lit fishing vessels in other regions.