Ensemble Machine Learning for Fish Abundance Prediction: A Multi-Model Stacking Approach with Environmental and Fisheries Data

Abstract

Effective fisheries management demands accurate forecasting of abundance to avoid excessive harvesting and ensuring marine biodiversity is maintained. Typical assessment techniques that mainly use previous catch data often have no predictive capacity to incorporate complicated, time-varying environmental factors. We tackle this problem with a new methodology that implements an ensemble Machine Learning technique. The use of a stacking ensemble is utilized, where four heterogeneous Level 0 base learners, Random Forest, XGBoost, K-Nearest Neighbors, and a Multi-Layer Perceptron, were classi- fied. A Ridge regression model is used as the Level 1 meta-learner to optimally combine the base learners’ predictions. The model was trained and validated using a dataset of 1000 observations collected through a timeframe of fifteen years (2010-2024) from the Manila Bay fisheries area. The features executed to forecast abundance include eleven environmental parameters including water temperature, pH, dissolved oxygen, and several nutrients combined with five important fisheries parameters including fish- ing area, gear type, and species. The combined features ensemble achieves excellent predictive performance with an R-squared of 0.9922, Mean Absolute Error of 37.35, and Root Mean Squared Error of 49.37. An ablation study demonstrated that combining all features significantly outperformed a strictly environment- only model, which returned an R-squared of negative 0.0045, and provided a more balanced prediction compared to the fisheries- only model (R-squared 0.9968). The resulting model offers ac- tionable harvest recommendations based on predicted abundance categories (Lean, Moderate, and Abundant), providing a robust tool for adaptive fisheries resource management.