Submitted to: Transactions of the ASAE
Publication Type: Peer reviewed journal
Publication Acceptance Date: 12/1/2004
Publication Date: 3/1/2005
Citation: Van Liew, M.W., Arnold, J.G., Bosch, D.D. 2005. Problems and potential of autocalibrating a hydrologic model. Transactions of the ASAE. 48(3):1025-1040. Interpretive Summary: Hydrologic simulation models are important tools for studying water supplies and pollution levels in streams and rivers. Models used for these kinds of studies contain parameters that describe watershed properties such as vegetative cover, soil characteristics, and landscape features. Parameters must be given watershed specific values to accurately simulate runoff from that watershed. This process of assigning appropriate values for a particular watershed is referred to as model calibration. The increasing complexity of simulation models has led to a significantly difficult task of model calibration that requires a great deal of human effort. Although an automated approach to calibration may offer a substantial savings in labor, the calibrated set of parameters may not necessarily provide acceptable model simulations for a particular watershed. To better understand the strengths and limitations of manual and automatic approaches to model calibration, a study was conducted to assess a newly developed autocalibration tool in the Soil and Water Assessment Tool (SWAT) watershed scale model. Performance of the model was tested on the Little River Experimental Watershed (LREW) in Georgia and the Little Washita River Experimental Watershed (LWREW) in Oklahoma, two USDA ARS research watersheds. A long record of streamflow data on each of the watersheds was used to test the manual approach and three automated calibration approaches. Results of the study show that for the LREW, the autocalibration approaches outperformed the manual calibration based on the overall fit of monthly and daily hydrographs. On the LWREW, the manual approach did a better job in estimating the monthly and daily hydrographs. For both watersheds, the manual approach generally outperformed the autocalibration methods based on the overall water balance. Findings from this study suggest that the autocalibration option in SWAT provides a labor saving tool that shows promising results. However, manual adjustments following autocalibration may be necessary to maintain the overall water balance and adequately represent the range of streamflow values. Caution should also be exercised in using the autocalibration tool so that calibrated values in the model are representative of watershed conditions.
Technical Abstract: An investigation was conducted to evaluate strengths and limitations of auto and manual calibration in the watershed scale model referred to as the Soil and Water Assessment Tool (SWAT). Performance of the model was tested on the Little River Experimental Watershed (LREW) in Georgia and the Little Washita River Experimental Watershed (LWREW) in Oklahoma, two USDA-ARS watersheds. A long record of multi-gage streamflow data on each of the watersheds was used for model calibration and validation. Model performance of the streamflow response in SWAT was assessed using a 6 parameter manual calibration based on mass balance and visual inspection of hydrographs and duration of daily flow curves, a 6 parameter autocalibration method based on the sum of squares of the residuals after ranking objective function (autoSSQR6), a 6 parameter method based on the sum of squares of residuals (SSQauto6), and an 11 parameter method based on the sum of square of residuals (SSQauto11). Results show that for both watersheds, the manual calibration generally outperformed the autocalibration methods based upon percent bias (PBIAS) and simulation of the range in magnitude of daily flows. For the calibration period on LREW subwatershed F, PBIAS was 0.0%, -24.0%, -21.5%, and +29.0% for the manual, SSQRauto6, SSQauto6, and SSQauto11 approaches, respectively. Based on the coefficient of efficiency, the SSQauto6 and SSQauto11 methods gave substantially better results than did the manual calibration on the LREW. On the LWREW, however, the manual approach did a better job estimating the coefficient of efficiency statistic. Results of this study suggest that for the practitioner who is faced with substantial time restrictions, the autocalibration option in SWAT provides a labor saving tool that shows promising results. However, manual adjustments following autocalibration may be necessary to maintain mass balance and adequately represent the range in magnitude of output variables. Caution should also be exercised in utilizing the autocalibration tool so that calibrated values in the model are representative of watershed conditions.