Advancing harmful algal bloom predictions using chlorophyll-a as an Indicator: Combining deep learning and EnKF data...

Area of Science:

• Environmental Science
• Data Science
• Oceanography

Background:

• Data-driven deep learning models are increasingly used for Harmful Algal Bloom (HABs) prediction.
• These models face limitations due to inherent structure and process uncertainties.
• Data assimilation (DA) offers a method to improve dynamic system predictions by integrating observations with model forecasts.

Purpose of the Study:

• To introduce and evaluate the application of data assimilation (DA) to enhance deep learning models for HABs prediction.
• To assess the impact of different assimilation frequencies on the accuracy of chlorophyll-a predictions.
• To determine the optimal frequency for incorporating new observations into HABs prediction models.

Main Methods:

• Developed 100 Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) models for one-day ahead chlorophyll-a prediction.
• Utilized high-frequency environmental data (pH, temperature, conductivity, turbidity, dissolved oxygen, ORP) as input.
• Employed an Ensemble Kalman Filter (EnKF) for assimilating chlorophyll-a observations and explored various assimilation frequencies.

Main Results:

• Data assimilation significantly improved the accuracy of chlorophyll-a predictions compared to models without DA.
• Daily assimilation yielded the lowest RMSE: 0.02 μg/l for LSTM and 0.03 μg/l for GRU.
• Monthly assimilation resulted in substantially higher RMSE: 3.59 μg/l for LSTM and 3.63 μg/l for GRU, indicating the importance of assimilation frequency.

Conclusions:

• Data assimilation is a viable strategy to enhance the accuracy and reliability of deep learning-based HABs monitoring models.
• Higher assimilation frequencies, particularly daily, lead to more precise chlorophyll-a predictions.
• Findings provide guidance on optimal data incorporation frequencies for HABs prediction models, supporting effective management.