Enhancing surface water quality prediction in data-scarce sites using transfer learning and neural networks

Abstract

Surface water quality prediction is essential for effective water treatment, pollution control, and regulatory compliance. However, neural network (NN)-based predictions are significantly constrained by the limited availability of high-quality training data at newly established or data-scarce monitoring stations. This study innovatively addresses this challenge by employing Transfer Learning (TL) to leverage existing knowledge from data-rich monitoring sites, aiming to improve predictive performance under conditions of data scarcity. To systematically enhance TL effectiveness, we developed and comprehensively evaluated six novel Similarity Measurement Indexes (SMIs) designed explicitly for optimal source domain selection. A case study involving five monitoring stations in Southern China demonstrated that the proposed TL methodology significantly improved prediction accuracy, achieving a substantial reduction of up to 79.9 % in RMSE compared with models trained solely on limited local data. Among the newly introduced SMIs, the P-RMSER and Distance indexes emerged as highly effective tools for identifying the most suitable source domains. Furthermore, we found that the selection of TL hyperparameters—particularly the number of frozen layers and fine-tuning learning rate—was critical in further optimizing predictive performance. These findings offer innovative practical guidelines and methodological advancements for achieving robust water quality forecasting in data-scarce environments.