Fishery knowledge-guided machine learning for spatial prediction of catch-per-unit-effort

Catch per unit effort (CPUE) is a key indicator of fish stock abundance. However, CPUE estimates derived from fishery logbooks are highly susceptible to noise and missing entries, leading to systematic bias in abundance estimation. To address this issue, we developed a comprehensive knowledge-guided machine learning (KGML) framework that incorporates both data preprocessing and model refinement, designed to enhance the accuracy and ecological consistency of spatial CPUE predictions. We applied this framework to neon flying squid ( Ommastrephes bartramii ) data collected in the Northwest Pacific during 2002–2019, using ocean environmental variables, including sea surface temperature, salinity, height, and chlorophyll-a, as factors for modeling and prediction. Guided by fishery expertise, we first constructed a refined dataset by removing implausible outliers and likely false-zero records. Initial experimental results confirmed that knowledge-guided data cleaning substantially improved model performance. However, subsequent Shapley additive explanations (SHAP) feature contribution analysis revealed spatial information dominated the feature importance rankings to an unreasonable degree, suggesting the model primarily memorized locations. To mitigate this effect, we further introduced a cost-aware loss function in model refinement, assigning a greater weight to the loss incurred by non-zero CPUE samples. The final SHAP analysis validated this refinement strategy, confirming a successful shift in the model's predictive focus from spatial memorization towards environmental drivers. In conclusion, this two-stage KGML approach not only maximized predictive accuracy and robustness but also significantly strengthened species distribution models by ensuring theoretical consistency in feature contributions. This provides a practical and robust framework for improving ecological indicators and supporting ecosystem-based fishery management, particularly in data-limited contexts. • Fishery knowledge-guided machine learning enhances spatial CPUE prediction fidelity. • Expert-informed data cleaning effectively removes false-zeros and reduces bias. • Cost-aware loss function mitigates bias by shifting focus to ecological drivers. • Extremely Randomized Trees achieve the most reliable predictive performance.