Enhanced prediction of river dissolved oxygen through feature- and model-based transfer learning

Area of Science:

• Environmental Science
• Data Science
• Hydrology

Background:

• Water quality monitoring data often exhibits non-uniformity across different locations within a basin.
• Extracting knowledge from data-rich sites to aid data-poor sites is crucial for effective water quality management.
• Transfer learning (TL) offers a promising approach to improve predictions at data-poor sites by leveraging data from data-rich sites.

Purpose of the Study:

• To compare and combine feature-based and model-based transfer learning methods for improving dissolved oxygen (DO) forecasting.
• To construct and evaluate Long Short-Term Memory (LSTM) models for water quality prediction using different TL strategies.
• To assess the effectiveness of combined TL approaches versus single-type TL for data-poor river sites.

Main Methods:

• Employed feature-based (Transfer Component Analysis - TCA) and model-based (pretraining and fine-tuning) transfer learning techniques.
• Developed Long Short-Term Memory (LSTM) models for dissolved oxygen (DO) forecasting in the West Channel of Guangzhou.
• Compared baseline LSTM models with models utilizing single-type and combined transfer learning strategies.

Main Results:

• The best single-type TL strategy (LSTM without TCA, freezing the fully connected layer after pretraining) improved 3-day DO prediction performance (Nash-Sutcliffe Efficiency - NSE) by up to 46.2% compared to the baseline.
• The best combined TL strategy (using TCA and freezing the second fully connected layer) further enhanced 3-day DO prediction performance (NSE) by up to 48.7% compared to the baseline.
• Combined feature- and model-based TL methods demonstrated superior DO prediction performance in data-poor river environments.

Conclusions:

• Combining feature-based and model-based transfer learning yields superior dissolved oxygen (DO) prediction performance in data-poor river systems.
• Transfer learning strategies significantly enhance the accuracy of LSTM models for water quality forecasting.
• This study provides valuable insights into optimizing TL approaches for environmental data analysis.