Deep learning decodes multi-horizon dynamics and probabilistic risks of harmful algal blooms

Area of Science:

• Environmental Science
• Data Science
• Ecology

Background:

• Harmful algal blooms (HABs) are increasing globally in freshwater systems due to climate change and eutrophication.
• Current predictive models for HABs lack hourly resolution, uncertainty quantification, and mechanistic interpretability.
• Existing methods hinder proactive intervention and mask ecological risks associated with high-frequency dynamics.

Purpose of the Study:

• To develop an interpretable deep neural network (DNN) for accurate, hourly resolution HAB forecasting.
• To integrate future environmental data with historical observations for improved prediction.
• To provide uncertainty-aware, mechanistically interpretable predictions for HABs.

Main Methods:

• Developed BloomNet, a novel DNN architecture for multi-horizon quantile forecasting.
• Integrated historical observations and future environmental covariates from a hyper-eutrophic lake.
• Evaluated BloomNet on a four-year hourly dataset, comparing it against LSTM, Transformer, and TCN models.

Main Results:

• BloomNet demonstrated exceptional predictive stability across 24-, 48-, and 72-hour horizons (R² up to 0.78).
• The model outperformed baseline deep learning methods, avoiding iterative error accumulation.
• Interpretability mechanisms revealed shifting ecological drivers (water temperature, total phosphorus) and predicted bloom onset 24 hours in advance.

Conclusions:

• BloomNet offers a significant advancement in HAB early warning systems, providing accurate and interpretable hourly predictions.
• The model's ability to quantify uncertainty and identify key drivers transforms water resource management towards a proactive approach.
• Probabilistic quantile outputs enable a graduated, risk-oriented warning protocol for enhanced ecological monitoring and intervention.