A new frontier in streamflow modeling in ungauged basins with sparse data: A modified generative adversarial network with explainable AI

Abstract

Streamflow forecasting is crucial for effective water resource planning and early warning systems, especially in regions with complex hydrological behaviors and uncertainties. While machine learning (ML) has gained popularity for streamflow prediction, many studies have overlooked the predictability of future events considering anthropogenic, static physiographic, and dynamic climate variables. This study, for the first time, used a modified generative adversarial network (GAN) based model to predict streamflow. The adversarial training concept modifies and enhances the existing data to embed featureful information enough to capture extreme events rather than generating synthetic data instances. The model was trained using (sparse data) a combination of anthropogenic, static physiographic, and dynamic climate variables obtained from an ungauged basin to predict monthly streamflow. The GAN-based model was interpreted for the first time using local interpretable model-agnostic explanations (LIME), explaining the decision-making process of the GAN-based model. The GAN-based model achieved R2 from 0.933 to 0.942 in training and 0.93–0.94 in testing. Also, the extreme events in the testing period have been reasonably well captured. The LIME explanations generally adhere to the physical explanations provided by related work. This approach looks promising as it worked well with sparse data from an ungauged basin. The authors suggest this approach for future research work that focuses on machine learning-based streamflow predictions.

Keywords

• Generative adversarial network
• Hydrology
• Machine learning
• Streamflow
• Explainable AI