Classification of set-net fish catch volumes in Iwate Prefecture, Japan using machine learning with water temperature and current distribution images at migration depth

Abstract

Efficient management of set-net fisheries can be achieved by predicting fish catches in advance. We aimed to estimate fish catch volumes using machine learning based on the data obtained from past set-net fisheries. We also suggested that the estimation can be conducted without the installation of special sensors, such as fish finders, sonars, and cameras, in the set nets. Therefore, a special feature of this study is the construction of a machine learning model using marine environmental images around the set net. In particular, the distribution images of water temperature and current at a certain water depth were applied to the model. In set-nets where migratory fish are caught, ocean data related to migration routes can be useful for estimation. The images used in this study included the water temperature at a depth of 100 m, absolute velocity at a depth of 50 m, and sea surface temperature for comparison. We then classified the fish catch into three classes as follows: chum salmon, yellowtail, and Japanese common squid in Iwate Prefecture, Japan. Consequently, our results achieved an accuracy of approximately 70–80% compared with the validation and test dataset with the water temperature at depths of 100 m. They showed higher accuracy than the sea surface temperature and absolute velocity at a depth of 50 m. This study indicates the usefulness of image-based classification for predicting the volumes of fish caught by set nets.