PHYSICS-INFORMED DEEP LEARNING FOR PREDICTING PHYTOPLANKTON DYNAMICS AND HYPOXIA IN ENCLOSED WATERS

Abstract

Accurate forecasting of phytoplankton dynamics and hypoxia is essential for effective management of freshwater ecosystems under increasing climatic and anthropogenic pressures. This study develops a hybrid deep learning framework that integrates physico-chemical mass-balance equations with neural network models (NN, RNN, and LSTM) to predict chlorophyll-a, -b, -c, and dissolved oxygen (DO) in Lake Biwa, Japan. Three approaches are examined: Case 1 using conventional observational inputs, Case 2 incorporating balance-equation–derived mechanistic inputs, and Case 3 further introducing physics-informed constraints in the loss function for DO prediction. Short-term forecasts show that Case 2 improves the reproduction of chlorophyll peaks at 0.5 m depth, particularly for chlorophyll-b, with LSTM reducing RMSE from 0.41 to 0.35 µg/L. For DO, Case 2 consistently reduces prediction errors at 90 m depth across all models by approximately 15–25%. Case 3 yields additional short-term error reductions, although without statistical significance. Paired t-tests confirm significant short-term improvements for DO in the NN model (p = 0.05) and for chlorophyll-b in the NN model (p = 0.02). In long-term forecasting, chlorophyll prediction accuracy decreases in Case 2 due to fixed hyperparameter settings, whereas DO predictions show statistically significant improvements in both Case 2 and Case 3 (p = 0.00), highlighting the effectiveness of physics-informed constraints for capturing slow oxygen dynamics. Overall, the proposed framework provides a process-informed basis for adaptive water quality management and hypoxia mitigation in freshwater lakes.