Causal-guided regional deep learning for chlorophyll‑a dynamics prediction in heterogeneous marine systems.

Harmful algal blooms pose increasing threats to marine ecosystems and human health, underscoring the need for reliable predictions of chlorophyll-a (Chl-a), a key indicator of algal biomass. However, existing deep learning models are often criticized as "black boxes," which limits ecological interpretability and reduces robustness under environmental regime shifts. In this study, we develop a causal-guided deep learning framework that integrates Liang-Kleeman information flow theory with convolutional long short-term memory (ConvLSTM) networks to predict Chl-a dynamics across the heterogeneous coastal and shelf seas of China. A causal analysis of ten environmental factors (e.g., nutrients, temperature, salinity, solar radiation, precipitation, currents) reveals pronounced regional heterogeneity in the drivers of Chl-a. Based on these causal structures, we partition the study area into three ecologically distinct zones: a complex, multi-driver region influenced by river plumes and Kuroshio intrusion (Region 1); an open-ocean region dominated by seasonal forcing (Region 2); and a shallow shelf region where advective processes play a notable role (Region 3). Region‑specific ConvLSTM models, using only causally validated core predictors, achieve excellent 1-day forecast accuracy (R² > 0.90 across all regions), comparable with full‑factor models. At a 15-day lead time, both models capture the seasonal signal but smooth daily variability, which is ecologically consistent for physically pulsed systems. Generally, the core-factor model outperforms the full-factor model across all regions, indicating that non-causal predictors introduce noise that degrades longer-term forecasts. These findings demonstrate that causally guided predictor selection enhances model simplicity, portability, and long-lead reliability. The proposed causal-zonation-ConvLSTM framework provides a transferable approach for region-specific water quality forecasting, with direct implications for eutrophication management and early warning of harmful algal blooms.