Uncertainty Analysis of Hydrologic and Water Quality Predictions for a Small Watershed Using SWAT2000

Abstract

Hydrologic and water quality (H/WQ) models are being used with increasing frequency to devise alternative pollution control strategies. It has been recognized that such models may have a large degree of uncertainty associated with their predictions, and that this uncertainty can significantly impact the utility of the model. In this study, ARRAMIS (Advanced Risk & Reliability Assessment Model) software package was used to analyze the uncertainty of the SWAT2000 (Soil and Water Assessment Tool) outputs concerning nutrients and sediment losses from agricultural lands. ARRAMIS applies Monte Carlo simulation technique connected with Latin hypercube sampling (LHS) scheme. This technique is applied to the Warner Creek watershed located in the Piedmont physiographic region of Maryland, and it provides an interval estimate of a range of values with an associated probability instead of a point estimate of a particular pollutant constituent. Uncertainty of model outputs was investigated using LHS scheme with restricted pairing for the model input sampling. Probability distribution functions (pdfs) for each of the 50 model simulations were constructed from these results. Model output distributions of interest in this analysis were stream flow, sediment, organic nitrogen (organic-N), organic phosphorus (organic-P), nitrate, ammonium, and mineral phosphorus (mineral-P) transported with water. Developed probability distribution functions for the model provided information with desirable probability. Results indicate that consideration of input parameter uncertainty produces 64% less mean stream flow along with approximately 8.2% larger sediment loading than obtained using mean input parameters. On the contrary, mean of outputs regarding nutrients such as nitrate, ammonia, organic-N, and organic-P (but not mineral-P) were almost the same as the one using mean input parameters. The uncertainty in predicted stream flow and sediment loading is large, but that for nutrient loadings is the same as that of the corresponding input parameters. This study concluded that using a best possible distribution for the input parameters to reflect the impact of soils and land use diversity in a small watershed on SWAT2000 model outputs may be more accurate than using average values for each input parameter.