Machine-learning and thresholding algorithms to automatically predict fishing effort of small-scale trawl fishery

Abstract

To assess fishery resources, it is necessary to easily obtain information on catch per unit effort, which is a resource indicator. In this study, two algorithms were developed for predicting the fishing effort (number of fishing operations, daily operating distance, and daily operating time) of a small-scale trawl fishery. These algorithms predict fishing efforts after preprocessing (including deleting outliers from the raw data), followed by classification of the operating conditions and threshold processing based on the operation period. One algorithm uses a machine-learning model for the classification process, and the other uses thresholding. The mean prediction error of the machine-learning algorithm on three datasets ranged from 1% to 11%, 2% to 8%, and 1% to 5% in terms of the number of operations, operating time, and operating distance, whereas that of the thresholding algorithm ranged from 3% to 52%, 2% to 5%, and 2% to 7%, respectively. A sensitivity analysis of the amount of training data indicated that prediction was possible using 5 days of training data. The developed algorithms are potentially useful for fish stock assessment.