Integrating advanced deep learning with ecological indicators for accurate prediction of harmful algal bloom occurrence and intensity

Harmful algal blooms (HABs) are intensifying globally under climate warming and eutrophication, yet reliable forecasting remains challenging due to nonlinear and highly correlated environmental drivers. This study develops a dual-task deep learning framework for HAB occurrence classification and bloom-intensity prediction using FT-Transformer, Neural Oblivious Decision Ensembles (NODE), and SAINT. To ensure data reliability, the dataset was audited to distinguish observed and GAN-generated records, and synthetic-data fidelity was validated through statistical, correlation, and visual analyses. Potential target leakage associated with Bloom_Index was explicitly assessed, and a no-Bloom_Index ablation experiment was conducted to evaluate model robustness. Ecological analysis revealed that HAB events were associated with elevated sea surface temperature, chlorophyll anomaly, and nitrogen enrichment, together with reduced dissolved oxygen. FT-Transformer achieved the best classification performance (accuracy = 0.995 ± 0.0032, F1-score = 0.985 ± 0.0091, AUROC = 0.998 ± 0.0036), while NODE provided the highest regression accuracy for Bloom_Index prediction (MAE = 0.0476 ± 0.0172; R² = 0.9316 ± 0.0231). Ablation results confirmed that predictive performance remained robust without Bloom_Index. SHAP analysis identified temperature, chlorophyll anomaly, nutrient dynamics, and oxygen depletion as dominant bloom drivers. The novelty of this study lies in integrating advanced deep tabular learning with synthetic-data validation, leakage-aware evaluation, and interpretable ecological attribution for robust HAB early-warning systems.