Bridging causality and deep learning for harmful algal bloom prediction

Accurate estimation of chlorophyll-a (Chl-a) is essential for monitoring harmful algal blooms (HABs), particularly in vulnerable coastal regions. However, most machine learning (ML) approaches rely on purely correlative patterns, often lacking causal interpretability and robustness under changing environmental conditions. This study introduces an enhanced causal machine learning framework that integrates causal discovery, treatment effect estimation, and deep learning within a Causally Informed Neural Network (CINN). Using 31 environmental predictors from MODIS, ERA5, and HYCOM over the Persian Gulf, a region where HABs threaten desalination, fisheries, and coastal ecosystems, the model embeds causal graphs derived from the DECI algorithm and average treatment effects from double machine learning. Monotonic causal constraints were incorporated to align predictions with ecological expectations. Results show that CINN and its monotonic extension (MCINN) consistently outperform baselines, including Random Forests, XGBoost, and Support Vector Machines, achieving R² up to 0.926 (a 10-17% improvement over baselines) while reducing RMSE by up to 25%. Mediation and sensitivity analyses confirm the causal validity of key drivers, including sea surface temperature, non-fluorescence line height, and nutrient fluxes. Uncertainty quantification and counterfactual simulations further demonstrate the framework's potential for operational early-warning systems and policy interventions. By bridging causality and deep learning, this framework delivers an interpretable, data-efficient, and uncertainty-aware solution for predicting algal blooms in data-scarce, climate-sensitive marine environments.