Short-term dissolved oxygen forecasting in lakes of the middle and lower Yangtze River basin using generative AI–enhanced machine learning

This study focused on two large lakes in the middle and lower Yangtze River basin, representing diverse geographical and environmental conditions. Machine learning (ML) models are widely applied in eco-environmental science, yet their performance is constrained in data-scarce scenarios such as newly constructed monitoring stations. This study presents a novel framework integrating generative AI (GAI) with ML algorithms to generate synthetic samples for training dataset augmentation, improving short-term dissolved oxygen (DO) prediction accuracy. Results demonstrate that Transformer trained on GAI-augmented datasets achieved substantial improvements over models trained solely on original observations. The optimal scheme improved MAE, RMSE, and R² by 31.40%, 35.16%, and 21.99% at HMS station, and by 24.55%, 26.42%, and 8.70% at GS station. Significant differences exist across stations in key features influencing DO prediction accuracy and optimal synthetic data ratios in training set. DO dynamics at HMS are primarily controlled by physical factors, while those at GS are governed by combined physical, biological, and chemical factors. A 60% synthetic data was suitable for HMS station, while 40% for GS station. Furthermore, integrating GAI models with ML models significantly enhances capability to capture extreme DO values. The proposed framework is modular and generalizable, enabling configuration of GAI and ML components based on lake-specific characteristics for optimal prediction. • Generative AI models effectively generate high-quality synthetic data from limited observations. • A GAI-enhanced ML framework significantly improves DO prediction at data-scarce stations. • Optimal synthetic data proportion in ML model training set is critical for prediction accuracy. • The proposed hybrid methodology effectively enhances the capture of extreme DO values.