Weakly-Supervised Keypoint Detector for Fishing Ground Estimation

Fishing ground estimation is an important task in the fishing industry. In this study, we focus on fisher's decision-making process based on sea surface temperature patterns, which we regard as a pattern recognition problem. Our method involves treating fishing grounds as keypoints and employing a deep learning-based keypoint detector to estimate their locations from sea surface temperature maps. However, the available amount of catch data, which is essential for annotation purposes, is insufficient. To tackle this issue, we utilize fishing vessel trajectory data as weak supervision, despite its limited information, owing to its abundance. Through experimental results, we demonstrated the superior performance of our method compared to approaches relying solely on catch data.