A Discussion Blog on "Knowledge-informed deep learning for hydrological model calibration: an application to Coal Creek Watershed in Colorado"
This paper, “Knowledge-informed deep learning for hydrological model calibration: an application to Coal Creek Watershed in Colorado,” mainly focuses on improving the efficiency and reducing the computational load in hydrological model calibration by identifying sensitive parameters to model responses and records that have some relations, thereby improving the efficiency and accuracy of the Advanced Terrestrial Simulator (ATS).
The study was conducted in the Coal Creek Watershed, which has an area of 53.2 km², an elevation ranges from 2706 m to 3770 m, and is a snow-process-dominated watershed. Mutual Information (MI) was carried out in two steps to exclude parameters that do not have any significant effect on the model outputs. The analysis centered on four major steps: first, a preliminary analysis was done to roughly identify the influencing parameters with fewer realizations; second, only the sensitive parameters to model outputs were selected by MI analysis with a large number of realizations; third, a knowledge-informed DL model was trained on the sensitive parameters and model responses; and fourth, the model accuracy was tested by estimating the parameters from observations (in three cases: Q only as response, ET only as response, and both Q and ET as model response). Prediction efficiency was calculated using two metrics: Nash–Sutcliffe efficiency and the modified Kling-Gupta efficiency.
The sensitivity analysis showed that discharge (Q) is controlled by subsurface permeability in the low-flow season, and during the high-flow season, the snow melting process was found to be most critical to the outlet discharge. Similarly, snow melt was found to be the most critical parameter in late autumn and winter, and transpiration was dominant to ET in the high-flow season. One interesting observation is that perm_s4 has an MI value of about 0.12 with Q on 30 realizations, which later decreases with an increase in the number of realizations. This suggests that the effect is local, not global.
The comparison between knowledge-informed inverse mapping and general inverse mapping showed that the knowledge-informed method outperforms the other, and the model performances were better with Q as the response variable compared to ET. In contrast, knowledge-informed DL deteriorated the model performance in the estimation of ET, which may be attributed to the uncertainty of the remote sensing product.
The multiyear analysis showed that single-year observations can yield similar calibration results to using multiple years at this watershed, but using dry-year observations increased the simulation of Q.
Some Points:
The results in this study are impressive as it is a snow-dominated watershed. It would be interesting to see what the output looks like if it is done for a watershed with diverse topography, like the Koshi Watershed of Nepal, which covers an elevation range of 70 meters to 8848 meters and has hilly, mountainous, and Terai regions and various types of land cover and geological formations.
In this study, MI-based parameter filtering uses fixed binning (10 bins). MI values can be sensitive to binning also, so if the results of multiple binning were computed, it would be clear, like it has been done for multiple realizations.