Fishing Ground Estimation using Deep-learning-based Keypoint Detector

In this paper, we propose a method for estimating fishing grounds using a deep-learning-based approach and marine environmental information. Ground-truth data, comprising fishing and non-fishing locations classified by fish catch, is used to train our deep-learning-based keypoint detector. This detector identifies fishing and non-fishing locations from 2D marine environment maps, such as sea surface temperature. Traditional keypoint detectors rely on fully annotated ground-truth data for training, meaning that all keypoints in the training data must be annotated. However, the fishing ground estimation can only utilize data from locations where fishing vessels have operated. This limitation in ground-truth data results in an under-detection of predictions when employing conventional L2 loss. To address this issue, we employ a new Hinge loss function, which improves the number of predictions. Our proposed model demonstrates significant improvements over the baseline model trained with L2 loss, both in terms of the number of predictions and evaluation metrics.