Sensitivity and Uncertainty Analysis for the Annual Phosphorus Loss Estimator Model

Abstract

Models are often used to predict phosphorus (P) loss from agricultural fields. Although it is commonly recognized that model predictions are inherently uncertain, few studies have addressed prediction uncertainties using P loss models. In this study we assessed the effect of model input error on predictions of annual P loss by the Annual P Loss Estimator (APLE) model. Our objectives were (i) to conduct a sensitivity analyses for all APLE input variables to determine which variables the model is most sensitive to, (ii) to determine whether the relatively easy-to-implement first-order approximation (FOA) method provides accurate estimates of model prediction uncertainties by comparing results with the more accurate Monte Carlo simulation (MCS) method, and (iii) to evaluate the performance of the APLE model against measured P loss data when uncertainties in model predictions and measured data are included. Our results showed that for low to moderate uncertainties in APLE input variables, the FOA method yields reasonable estimates of model prediction uncertainties, although for cases where manure solid content is between 14 and 17%, the FOA method may not be as accurate as the MCS method due to a discontinuity in the manure P loss component of APLE at a manure solid content of 15%. The estimated uncertainties in APLE predictions based on assumed errors in the input variables ranged from ±2 to 64% of the predicted value. Results from this study highlight the importance of including reasonable estimates of model uncertainty when using models to predict P loss.

Improved APEX Model Simulation of Buffer Water Quality Benefits at Field Scale

“…For example, Moriasi et al (2016) determined the uncertainty of APEX-simulated annual runoff to be ±3% for a small watershed in Ohio. Bolster et al (2014) estimated the uncertainty of P predictions by the Annual P Loss Estimator (APLE) model (Vadas et al, 2009) to be within 28% for a high uncertainty of input parameters. Thus, the differences between buffer effectiveness calculated with measured or simulated data (up to 8% for runoff and 3% for TP) were within the same range as or smaller than common uncertainty ranges associated with measured data and modeling results.…”

Using AnnAGNPS to Predict the Effects of Tile Drainage Control on Nutrient and Sediment Loads for a River Basin

“…The sensitivity analyses approach used in this study follows those used in other AnnAGNPS modeling studies (Yuan et al, 2003;Das et al, 2008;Liu et al, 2008;Parker et al, 2008), with due recognition of some of the limitations of the one-at-a-time sensitivity measures for nonlinear models (Bolster and Vadas, 2013). The sensitivity approach taken herein was performed principally to descriptively document consistencies, or the lack thereof, with sensitive parameters in the literature to justify variable selection for model calibration/validation.…”

Modelling Phosphorus Losses from Tropical Agricultural Soils in Gilgel Gibe Watershed, Ethiopia

“…Sensitivity and uncertainty analysis done for APLE model using First Order Approximation (FOA) and Monte Carol Simulation (MCS) showed that the MCS was found better for its overall uncertainty analysis [21]. Therefore, in this study the MCS technique was employed after reducing the DP man component of equation 2 due to the absence of manure application on the study area.…”

“…To make the simulated output credible and the model prediction valid, the uncertainty analyses were performed using Monte Carlo simulation with an adequate number of simulations (i.e. 200000) by assuming a triangular distribution of uncertainty which gives better distribution of data sets [21]. The model prediction uncertainties were calculated using the same input data sets as of sensitivity analysis ( Table 5).…”

“…Relatively lower uncertainties can occur due to the remaining parameters. Uncertainties in measured P loss data are a function of errors introduced when measuring runoff, erosion, and concentration of P in solution and attached to sediment [21]. After Monte Carlo simulation, the predicted value of P loss to 95% CI [10.59, 10.94] was10.77 ± 0.17 kg ha -1 .…”