Knowing where and when fish are likely to be caught is central to sustainable fisheries management: it enables reduced emissions, bycatch, and ecosystem impacts while improving profitability and crew safety. Yet spatial predictions of catch rates across large marine regions remain coarse, infrequent, and often limited to single species or gear types. This paper addresses that gap by combining detailed vessel activity records, oceanographic and geospatial data, and modern deep-learning methods to generate high-resolution predictions of catch per unit effort (CPUE) across the Norwegian and Barents Seas.
The predictive framework uses a fully convolutional neural network that ingests thirty input channels describing each grid cell at each time step, aggregated over vessel visits. Inputs include static seafloor attributes (depth, slope, sediment composition) and dynamic oceanographic variables (temperature profiles, salinity, currents, chlorophyll). Positional and seasonal encodings are included so the model can learn geographic and temporal dependencies—recognizing that the same environmental measurements may imply different ecological outcomes depending on location and season.
The empirical case study focuses on cod, saithe, and haddock, each caught with bottom trawl and Danish seine. Predictions are produced on a high-resolution spatial and temporal grid spanning the Norwegian coast to the high Arctic, at weekly-to-monthly time steps for 2011–2025. Ground-truth CPUE estimates are constructed by integrating mandatory electronic catch reports with satellite-based vessel-tracking records, allowing estimation of both total catch and the spatial distribution of fishing effort (CPUE in kilograms). The model is trained on 2012–2021 data, validated on 2022–2023, and evaluated on held-out 2024–2025 observations.
Planned extensions include incorporating annual stock-assessment indicators (spawning biomass, recruitment, fishing mortality) to help separate true stock abundance from spatially concentrated fishing pressure, and developing a hierarchical model that separates broad interannual stock trends from fine-scale environmental responses. We are also exploring uncertainty quantification methods so predictions in data-sparse regions or anomalous periods can be reported with explicit confidence intervals rather than point estimates alone. The long-term goal is a system that supports operational fisheries planning and advances scientific understanding of how fish distributions respond to environmental variability and climate-driven ocean change.
 

暂无