This paper proposes a Positive Unlabeled (PU) learning approach to narrow down the prediction area in the context of good fishing ground prediction. PU learning, a type of semi-supervised learning, is particularly suitable for scenarios characterized by limited positive examples and abundant unlabeled data, as often encountered in fishing ground prediction tasks, where the explored sea areas identified as productive fishing spots are treated as positive instances and the vast unexplored sea areas as unlabeled data. Conventional methods often struggle to accurately model the characteristics of good fishing grounds, resulting in overly broad prediction areas or overly restrictive constraints. To tackle this challenge, we present a PU learning-based method designed to identify negative examples from the unlabeled data and consequently refine the area predicted as positive. Specifically, we train a prediction model for fishing duration, which can be considered a surrogate indicator of good fishing grounds. Subsequently, we apply this model to predict the fishing duration for unexplored areas; those areas exhibiting short fishing durations are deemed reliable negative examples. By incorporating both past positive examples and selected negative examples into a binary classification framework for predicting good fishing grounds, we aim to fine-tune the prediction area. To the best of our knowledge, this study represents the first application of PU learning in the domain of good fishing ground prediction. Experimental comparisons conducted using data from bullet tuna trolling validate the efficacy of the proposed methodology.