Information support for chub mackerel Scomber japonicus fishery in the Pacific waters of the Russian Federation

Abstract

Methods of machine learning were applied for forecasting of chub mackerel fishing grounds in the South Kuril fishery district. The problem of perspective fishing area definition was reduced for a binary classification task, i.e. the sets of environmental conditions corresponded with presence or absence of fishing operations were determined for each point within the district. The fishery statistics for 2016–2020 and the data on SST with delay of 4–7 days from the date of catch, spatial SST gradients calculated using Belkin algorithm, and day-to-day SST variations were processed using LightGBM machine learning algorithm. The model was trained on the data for 2016–2019 and verified on the data for 2020. The AUC (as an aggregate measure of performance across all possible classification thresholds) varied from 0.65 to 0.92. In the fishery season of 2020, AUC was 0.69, on average, growing to 0.75 in the period of the highest catches. Approximately 75 % of the annual catch of chub mackerel was caught at the predicted sites in 2020; this portion reached 84 % in the period of the highest catches.

Keywords

• chub mackerel
• scomber japonicus
• pelagic fishery season
• south kuril fishery district
• operational forecasting
• machine learning