| Summary: | A Bayesian-Monte Carlo approach was carried out to assess uncertainties in process-based, continuous simulation models. This was achieved by using the 93.13 version of the Water Erosion Prediction Project (WEPP) model when applied to a small semi-arid rangeland watershed nested in the Walnut Gulch Experimental Watershed, near Tombstone, AZ. Two techniques were evaluated to calibrate the model and identify the probability distributions of parameters based on the concept of model output classification ("acceptable" or "not acceptable"). Technique I consisted of Monte Carlo simulation with correlated parameter deviates generation. Technique II applied Monte Carlo simulation with correlated parameter deviates generation within a Bayesian framework to update parameter probability distributions every time that the model produced an acceptable realization. Based on the results, both techniques were able to calibrate the model and to identify parameter distributions, however; Technique I was computational more expensive than Technique II. This resulted because Technique II searched for parameter deviates within the region of the prior distributions more likely to produce acceptable model realizations. The contribution of parameter error and model error to total model uncertainty was assessed by using the mean square error equation. Errors were uniform during continuous simulations, errors never increased or decreased with the time of simulation. However, errors are larger toward components of higher levels of aggregation (soil erosion calculations). This resulted in larger errors in sediment yield predictions. Lack of homoscedasticity was observed, the largest errors for the largest rainfall events. This is more evident for peak runoff and sediment yield than for runoff volume. Also, a larger contribution of model error to total prediction uncertainty for peak runoff and sediment yield predictions was observed. Prediction intervals of runoff volume indicated that WEPP does acceptable responses in estimating infiltration variables. Almost all observed runoff volume data were inside the 90% prediction intervals. Prediction intervals for peak runoff revealed that WEPP rarely comprised the observed data within the range of predictions. Because the large errors in estimating sediment yield, most of the observed data never fell inside the prediction intervals.
|
|---|
