Benchmarking deep learning model estimates of hydrologic pathway contribution to streams and springs using conceptual and physically-based models

Hydrologic models are powerful tools for the assessment and management of water resources. One particular quantity of interest is the relative fraction of discharge derived from runoff, interflow, and baseflow. Over the last half-century, process-driven models - those developed to serve as proxies for physical phenomena - have been the prevailing tool used by modelers to make these types of estimates. However, recent advancements in data-driven models have yielded improved hydrologic prediction capabilities, but often at the cost of physical interpretability. There is potential for benchmarking hydrologic pathway estimates derived from data-driven models against those from process-based models to achieve the benefit of improved prediction as well as the extraction of physical-analogues. In this study, we use physically-based (SWAT), conceptually-based (LUMP), and deep-learning (LSTM) models to simulate hydrologic pathway contributions for a fluvial stream and a karst spring. Runoff, interflow, and baseflow estimates from the LSTM model were generated using a recursive digital filter (RDF) and benchmarked against process-driven estimates. Results showed that while all models performed satisfactorily, the LSTM model outperformed both the SWAT and LUMP models in estimating stream and spring discharge, respectively. This improved performance was achieved with only 25% of the observed time-series data used for training, a condition imposed on the data-driven model to provide a like-for-like comparison to calibration periods of the process-driven models. Regarding pathways, the LSTM+RDF model was able to successfully match the magnitude of process-based estimates of runoff, interflow, and baseflow (ρ ranging from 0.58 to 0.71). However, the process-based models exhibited more realistic time-fractal scaling of hydrologic pathways than did the LSTM+RDF model. We demonstrate the utility and potential limitations of extracting physical-analogues from LSTM modeling, which will be useful as deep learning approaches to hydrologic modeling become more prominent.