Simultaneous feature engineering and interpretation: Forecasting harmful algal blooms using a deep learning approach.

Area of Science:

• Environmental Science
• Water Quality Management
• Computational Hydrology

Background:

• Routine monitoring for harmful algal blooms (HABs) often uses low temporal frequency, insufficient for dynamic cyanobacteria variations.
• Accurate forecasting of HABs is crucial for water quality and resource management.

Purpose of the Study:

• To develop and evaluate a deep learning model, RETAIN-D, for high-resolution forecasting of cyanobacteria abundance.
• To improve temporal resolution, forecasting performance, and interpretability in HABs prediction.

Main Methods:

• Developed RETAIN-D, a model incorporating reverse time attention with a decay mechanism.
• Applied RETAIN-D to forecast cyanobacteria abundance in the Nakdong and Yeongsan rivers, South Korea.
• Compared RETAIN-D with LSTM, GRU, and RETAIN models using meteorological, hydrological, environmental, and biological data.

Main Results:

• RETAIN-D achieved high accuracy in forecasting cyanobacteria abundance (e.g., R² = 0.76-0.98) on a daily resolution.
• RETAIN-D outperformed LSTM, GRU, and RETAIN models in forecasting accuracy across sites and outputs.
• Ambient temperature, irradiance, flow rates, and residence time were identified as key predictive features, with seasonal variations.

Conclusions:

• RETAIN-D offers a significant advancement in HABs forecasting, providing high temporal resolution and accuracy.
• The model's interpretability aids in understanding factors influencing cyanobacteria blooms.
• RETAIN-D is a versatile tool applicable to various forecasting models requiring enhanced resolution, performance, and interpretability.