Algal bloom forecasting with time-frequency analysis: A hybrid deep learning approach.

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Algal blooms pose significant environmental challenges.
• Traditional forecasting methods struggle with the complex dynamics of algal growth.
• Deep learning, specifically Long-Short-Term Memory (LSTM) networks, shows promise but is limited by data non-stationarity.

Purpose of the Study:

• To enhance the accuracy and adaptability of LSTM models for algal bloom forecasting.
• To introduce a novel hybrid approach combining Wavelet Analysis (WA) with LSTM.
• To validate the performance of the proposed WLSTM model against existing methods.

Main Methods:

• Developed a hybrid Wavelet-LSTM (WLSTM) model.
• Applied time-frequency Wavelet Analysis (WA) to preprocess data for LSTM.
• Evaluated WLSTM on Lake Mendota (Wisconsin) and Lake Tuesday (Michigan) datasets.

Main Results:

• WLSTM reduced forecasting inaccuracy by 41% ± 8% compared to classic LSTMs.
• Achieved high prediction accuracy (R² = 0.976, 0.878, 0.814) at hourly, daily, and monthly resolutions.
• Outperformed Deep Neural Network (DNN) and ARIMA models, reducing RMSE by 72% and 85%, respectively.
• Successfully forecasted extreme algal bloom events with >95% accuracy.

Conclusions:

• The WLSTM model offers a significant advancement in algal bloom forecasting.
• Hybrid deep learning approaches can effectively address aquatic environmental challenges.
• This work provides a robust tool for early warning systems and environmental management.