Knowledge-Informed Deep Learning for Hydrological Model Calibration: An Application to Coal Creek Watershed in Colorado
Abstract. Deep learning (DL)-assisted inverse mapping has shown promise in hydrological model calibration by directly estimating parameters from observations. However, the increasing computational demand for running the state-of-the-art hydrological model limits sufficient ensemble runs for its calibration. In this work, we present a novel knowledge-informed deep learning method that can efficiently conduct the calibration using a few hundred realizations. The method involves two steps. First, we determine decisive model parameters from a complete parameter set based on the mutual information (MI) between model responses and each parameter computed by a limited number of realizations (~50). Second, we perform more ensemble runs (e.g., several hundred) to generate the training sets for the inverse mapping, which selects informative model responses for estimating each parameter using MI-based parameter sensitivity. We applied this new DL-based method to calibrate a process-based integrated hydrological model, the Advanced Terrestrial Simulator (ATS), at Coal Creek Watershed, CO. The calibration is performed against observed stream discharge (Q) and remotely sensed evapotranspiration (ET) from the water year 2016 to 2018. Preliminary MI analysis on 50 realizations resulted in a down-selection of seven out of fourteen ATS model parameters. Then, we performed a complete MI analysis on 396 realizations and constructed the inverse mapping from informative responses to each of the selected parameters using a deep neural network. Compared with calibration using all observations, the new inverse mapping improves parameter estimations, thus enhancing the performance of ATS forward model runs. The Nash-Sutcliffe efficiency (NSE) of streamflow predictions increases from 0.65 to 0.80 when calibrating against Q alone. Using ET observation, on the other hand, does not show much improvement on the performance of ATS modeling mainly due to both the uncertainty of the remotely sensed product and the insufficient coverage of the model ET ensemble in capturing the observation. By using observed Q only, we further performed a multi-year analysis and show that Q is best simulated (NSE: 0.85) by including in calibration the dry year flow dynamics that shows more sensitivity to subsurface characteristics than the other wet years. Our success highlights the importance of leveraging data-driven knowledge in DL-assisted hydrological model calibration.
Peishi Jiang et al.
Status: final response (author comments only)
RC1: 'Comment on hess-2022-282', Anonymous Referee #1, 19 Sep 2022
- AC1: 'Reply on RC1', Peishi Jiang, 28 Dec 2022
RC2: 'Comment on hess-2022-282', Anonymous Referee #2, 10 Oct 2022
- AC2: 'Reply on RC2', Peishi Jiang, 28 Dec 2022
RC3: 'Comment on hess-2022-282', Anonymous Referee #3, 04 Dec 2022
- AC3: 'Reply on RC3', Peishi Jiang, 28 Dec 2022
Peishi Jiang et al.
Peishi Jiang et al.
Viewed (geographical distribution)
This study showcases a deep learning optimization method for a high-resolution hydrologic model supported by information theory. I appreciate the honest evaluation of the methodology, in-depth reasoning of the deteriorating model performance for ET, and examination of results and conclusions aligned with earlier studies. In general, this paper is well-written with a novel contribution. However, I think the paper would be stronger if the authors can address the following comments.
L158: Can the authors elaborate on what five soil types and four geological types are?
L160: A 1000-year spin-up is extremely long. Can the authors briefly explain the reason for this long spin-up even if it might be explained in Shuai et al 2022?
L162: Could the authors briefly explain how they preselected the parameters in this study?
L208: Does the MI have to be zero? If the MI between a parameter and the model responses is small enough, is it possible to neglect that parameter? What would be a proper threshold for it?
L208-210: Interesting! Great summary!
L215: When training using different combinations of years, why do the authors only look at Q, not ET?
L249-250: Given the narrowed list, it seems that the authors eliminated the parameters with small MI (not zero), which slightly contradicts the previous statement where only parameters with zero MI would be eliminated (L208). It would be helpful to clarify the threshold of MI below which the parameters will be eliminated.
L286-287: Please clarify whether the extrapolation issue partially or solely contributes to the worse MI-informed results.
L320: Very interesting results!
Author name: Should the third author be Alexander?