Deep learning to enhance and accelerate watershed model calibration
Abstract
Watershed simulation models such as the Soil and Water Assessment Tool (SWAT) consist of effective/conceptual parameters, which are difficult to measure directly. Also, models such as SWAT have many parameters that need to be calibrated. Hence, observational data such as streamflow are used to calibrate SWAT model parameters. Various techniques and software exist in literature to calibrate watershed models. However, existing calibration methods usually require reliable prior information and can be computationally expensive. Therefore, using current methods can result in reduced accuracy when calibrating high-dimensional SWAT parameters. To overcome these challenges, we present a fast and accurate parameter estimation workflow using deep learning (DL) to calibrate watershed models better than traditional methods. A significant advantage of the proposed DL workflow is that it accurately estimates sensitive parameters even under a high observation noise level (e.g., 25% relative observational errors). Moreover, these estimated parameters are well clustered (i.e., deviation in estimated parameters is minimal), which shows the robustness of the proposed workflow to observational errors. Compared to the Generalized likelihood uncertainty estimation (GLUE) method, the parameters estimated by the DL-enabled inverse models provide more accurate predictions of streamflow within and beyond the calibration period. The best DL-based calibrated set has an R2, Nash-Sutcliffe efficiency (NSE), logarithmic Nash-Sutcliffe efficiency (logNSE), and Kling-Gupta efficiency (KGE) scores to be 0.53, 0.67, 0.78, and 0.74, respectively. The best GLUE-based calibrated set has R2, NSE, logNSE, and KGE scores to be 0.48, 0.6, 0.7, and 0.68, respectively. The scores mentioned above show that the DL calibration set provides more accurate predictions of low and high streamflow. Another advantage of the proposed DL-based inverse models is that it is at least 103 times faster than the GLUE-based method. Due to the savings in computational cost, our DL-enabled parameter estimation is ideal for calibrating complex watershed models at larger scales.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H32B..09M