LSTM networks provide efficient cyanobacterial blooms forecasting even with incomplete spatio-temporal data

Area of Science:

• Environmental Science
• Ecology
• Water Quality Management

Background:

• Cyanobacterial blooms are increasing in inland waters due to human activities and climate change.
• These blooms threaten ecosystem health and water quality, especially toxin-producing strains.
• Early Warning Systems (EWS) are crucial for timely management of cyanobacterial blooms.

Purpose of the Study:

• To develop and evaluate an effective EWS for forecasting cyanobacterial bloom development.
• To compare the performance of six different predictive models using spatio-temporal data.
• To assess forecasting accuracy across multiple time horizons and using a hybrid evaluation system.

Main Methods:

• Utilized 6 years of incomplete, high-frequency spatio-temporal data from multiparametric probes, focusing on phycocyanin (PC) fluorescence.
• Developed a probe-agnostic method for data pre-processing and time series generation for bloom forecasting.
• Compared six predictive models (Linear Regression, Random Forest, LSTM - autoregressive and multivariate) using regression, classification, and skill metrics.

Main Results:

• The multivariate Long-Term Short-Term (LSTM) neural network demonstrated the best and most consistent performance across all forecasting horizons and metrics.
• LSTM achieved up to 90% accuracy in predicting a proposed PC alarm level (10 µg PC/L).
• Positive skill values confirmed LSTM's effectiveness in forecasting cyanobacterial blooms 16 to 28 days in advance.

Conclusions:

• The proposed EWS, particularly using multivariate LSTM, is highly effective for forecasting cyanobacterial blooms.
• This system offers significant advance warning, enabling proactive management strategies.
• The methodology provides a replicable approach for monitoring and predicting harmful algal blooms in inland waters.