Transfer learning for neural network model in chlorophyll-a dynamics prediction.

Area of Science:

• Environmental Science
• Data Science
• Ecological Modeling

Background:

• Neural network models predict chlorophyll-a dynamics but suffer from decreasing generalization ability over time.
• Model performance and accuracy degrade, necessitating improved prediction strategies.

Purpose of the Study:

• To optimize neural network models (FNN, RNN, LSTM) for chlorophyll-a prediction using transfer learning (TL).
• To enhance model generalization ability and maintain long-term prediction accuracy.
• To compare TL with parameter norm penalties (PNP) and dropout for improving generalization.

Main Methods:

• Employed transfer learning (TL) to optimize feedforward neural networks (FNN), recurrent neural networks (RNN), and long short-term memory (LSTM) models.
• Applied models to predict chlorophyll-a dynamics at 5-min intervals in an eastern China estuary reservoir.
• Compared TL with parameter norm penalties (PNP) and dropout for model generalization.

Main Results:

• Transfer learning (TL) models demonstrated superior prediction results compared to original models, PNP, and dropout.
• TL models significantly improved accuracy and maintained high performance in long-term applications.
• Original models and those with PNP/dropout lost predictive ability within 3 months, unlike TL models.

Conclusions:

• Transfer learning (TL) is an effective method for optimizing neural networks to predict chlorophyll-a concentration.
• TL enhances model generalization, reduces data distribution influence, and ensures sustained high accuracy for environmental monitoring.
• TL offers a significant advantage over PNP and dropout for long-term, accurate ecological predictions.