Interpreting the Effects of Environmental Variables on a Multistep Deep Learning Model for Algal Bloom Prediction...

Area of Science:

• Environmental science
• Machine learning
• Artificial intelligence

Background:

• Algal blooms pose significant environmental and economic challenges.
• Accurate prediction of chlorophyll-a concentration is vital for monitoring algal blooms.
• Machine learning offers potential for improved predictive capabilities in environmental monitoring.

Purpose of the Study:

• To develop and evaluate a sequence-to-sequence (Seq2Seq) deep learning model for predicting chlorophyll-a concentration.
• To analyze the impact of varying prediction intervals (1 to 28 days) on model performance.
• To investigate the influence of environmental variables on prediction accuracy using explainable artificial intelligence (XAI).

Main Methods:

• Development of a Seq2Seq deep learning model for chlorophyll-a concentration prediction.
• Evaluation of model performance across eight distinct time steps (t+1 to t+28 days).
• Application of Shapley additive explanations (XAI) to interpret model predictions and variable importance.

Main Results:

• The one-step-ahead prediction model achieved the highest performance (Nash-Sutcliffe efficiency [NSE] = 0.908).
• Model performance decreased significantly with longer prediction intervals, with NSE dropping to 0.255 for 28-day predictions.
• Flow rate was more important for short-term predictions, while sunshine duration became more influential for long-term predictions.

Conclusions:

• Deep learning models can effectively predict chlorophyll-a concentration, but performance is interval-dependent.
• XAI methods provide valuable insights into the factors driving prediction accuracy across different time scales.
• This research enhances the practical application of machine learning in environmental monitoring and algal bloom prediction.