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

Water quality monitoring data from various points within the same basin often show non-uniformity. A key scientific question is how to extract relevant knowledge from data-rich sites (source domains) and leverage the possible inter-site consistency of water quality to compensate for the limitations of data-poor sites (target domains). Transfer learning (TL) methods can improve the applicability of water quality predictions for data-poor sites but their comparison and combination have not been fully explored. This study employs feature-based (Transfer Component Analysis, TCA) and model-based (pretraining and fine-tuning) transfer learning, to assist in constructing Long Short-Term Memory (LSTM) models for forecasting the dissolved oxygen (DO) levels in the West Channel of Guangzhou, southern coastal China. The LSTM models at Yagang and Shimen stations were constructed as the basic and baseline models for source and target domains, respectively. By comparing and selecting different transfer learning strategies, the best single-type TL strategy emerged as a multi-sequence LSTM model without TCA but with the fully connected layer frozen after pretraining. It achieved increases in validation Nash efficiency coefficient (NSE) of 5.2%, 10.8%, and 46.2% for predicting DO over the next 3 days, respectively, compared to the baseline LSTM model at Shimen station. The best combined TL strategy involved using TCA and freezing the second fully connected layer in a multi-sequence LSTM model. It improved upon the baseline LSTM model with a validation NSE increase of 5.3%, 21.4%, and 48.7% over the next three days, respectively. This study demonstrates that combining feature- and model-based transfer learning methods can yield better DO prediction performance in data-poor rivers than using a single-type transfer learning method.