Forecasting Primary Productivity and Tuna Fishery Biomass in the North Pacific Subtropical Gyre with Machine Learning

Abstract

Effective fisheries management requires timely, quantitative guidance on evolving ecosystem conditions to balance stock sustainability with economic viability. This study addresses the need for proactive tools by developing a machine learning framework to improve the short- to medium-term predictability of Net Primary Production (NPP) and fish stock health within the North Pacific Subtropical Gyre. By leveraging in situ observations from the Hawaiian Ocean Time-series (HOT) and regional fishery datasets, the project aimed to translate biological and environmental measurements into actionable management forecasts. The analysis integrated diverse data sources, including ship-based NPP and hydrographic observations from Station ALOHA (1988–2023), climate indices (ENSO and PDO), and regional stock assessments for three tuna species: albacore, yellowfin, and bigeye (1952 – 2021). A suite of ML models, including linear regularized regressions, tree-based models, and neural networks, were trained to predict NPP at horizons between one and 24 months and fish biomass one and two years ahead. Model performance for both NPP and fish biomass was consistently poor across all forecast horizons, with negative R² values indicating that none of the models reliably captured variance beyond a naïve baseline. The poor results likely stem from the loss of granular signal during data aggregation and reconstruction. While some tracked the direction and magnitude of variability, the overall lack of predictive skill indicates that the current feature sets and data configurations are insufficient to capture complex ecosystem dynamics at the resolution required for management use. These results are nonetheless diagnostic. The failure was consistent across architectures, pointing to data constraints rather than modeling choices as the binding limitation. NPP models with sufficient temporal resolution to capture episodic events, combined with mechanistic features encoding trophic energy transfer and biophysical growth relationships, represent the most promising path forward. Combining satellite missions could match the present observational timeline and reduce dependence on reconstructed data. The framework developed here provides a replicable structure for climate-informed fisheries forecasting. Once refined with higher-quality data, it could support dynamic catch and effort recommendations that explicitly account for bottom-up environmental forcing on fish biomass, moving management beyond reliance on historical recruitment statistics alone. This study establishes the groundwork, and the next step is supplying it with the data it needs to perform and be optimized.