Evaluating Rainfall-Runoff Generation Mechanisms of Deep Learning Models Using a Process-Based Rainfall-Runoff Model

Deep learning models, particularly Long Short-Term Memory (LSTM) neural networks, have emerged as promising tools for streamflow simulation. While current research predominantly focuses on evaluating model performance in simulation tasks or through explainable machine learning approaches, studies examining whether LSTMs can accurately represent the physical processes of the hydrological cycle remain limited. This study assessed the LSTM’s capability to describe the physical processes in rainfall-runoff relationships through synthetic experiments. The core of this research relies on ideal hydrometeorological data and rainfall-runoff processes that can be represented through the Soil and Water Assessment Tool (SWAT) model. The LSTMs were trained using simulated streamflow from SWAT to avoid the unknown uncertainties inherent in observational data. Our results demonstrated that although the trained LSTM can perform well in streamflow simulation, with a Nash-Sutcliffe Efficiency of greater than 0.91. The rainfall-runoff lag times from LSTM models resemble those from the SWAT model. Despite that, the LSTM model fails a simple mass balance test, producing streamflow without rainfall. Runoff generation in LSTM is driven by correlation rather than causation. Temperature is an important variable for runoff simulation using the SWAT model, but it has minimal effects on runoff simulation with LSTM models. Results from our study underscore the need to understand rainfall-runoff mechanisms in LSTM models before applying them to unseen data or new environmental conditions.