Estimation of the spatiotemporal distribution of fish and fishing grounds from surveillance information using machine learning: The case of short mackerel (Rastrelliger brachysoma) in the Andaman Sea, Thailand

Abstract

Rapid and accurate data gathering on fish catches and the geographical distribution of fishing activities is essential in fish stock assessments and fisheries management. Surveillance systems like the Vessel Monitoring System (VMS) have become popular globally, providing accurate data on fishing activities and fish abundance in near real-time. Our study demonstrates a procedure to find the best method in machine learning algorithms for detecting fishing operations from surveillance data. VMS data of 3,237 trips conducted by 91 purse seiners targeting short mackerel (Rastrelliger brachysoma) in Thai waters of the Andaman Sea during 2020 were used for the study. Twenty-five spatiotemporal, environmental, and fisheries-related parameters were investigated as predictors to detect and map the fishing activity. Eight algorithms were used to compare the model performance using cross-validation. The total catch per trip was reallocated to the predicted fishing locations of the trip from the best algorithms to depict the fishing ground. The results show that operation time and vessel speed are essential elements for determining the fishing operation of purse seiners. The most-relevant predictors are the calculated speed, operation time, and instantaneous speed. The Random Forest (RF) algorithm performed the best, with 86% accuracy, and therefore was used as the final model to detect the fishing operations. However, all reallocation approaches performed similarly well in depicting fish distribution. The allocation of catch data equally to the fishing position is recommended since it is the most straightforward technique. In contrast, reallocation catch using the RF algorithm prediction ratio for visualising the CPUE distribution achieved the best performance (with a slope of 1) compared with the training data. This study provides the first decent procedures for using surveillance information to detect and identify the prospective fishing operation of purse seiners—an approach that allows for a better understanding of short mackerel distribution. The proposed procedures should help with time–area-based management of intensively exploited migratory fish species, which requires rapid and accurate information.