Comparison of Three Spatial Sensitivity Analsis Techniques
Nathalie Saint-Geours1 and Linda Lilburne2
1.AgroParisTech, UMR TETIS, F-34035, Montpellier, France
2.Landcare Research, PO Box 40, Lincoln 7640, Canterbury, New Zealand
firstname.lastname@example.org; 2. email@example.com
Abstract: This paper compares the spatial Sobol' sensitivity analysis approach to two other sensitivity analysis techniques on a model with spatially distributed inputs. The comparison is performed on AquiferSim, a model that simulates groundwater flow and nitrate transport from paddock (i.e. field) to aquifer. Some of the input layers have considerable uncertainty. Alternative soil and land-use layers were simulated through Monte Carlo simulation based on expert-derived confusion matrices. Uncertainty of the raster rainfall layer was simulated via geostatistical unconditional simulation of error fields. The three sensitivity techniques are: (1) the spatial Sobol' technique, (2) one-at-a- time (OAT) variation around base sample points, and (3) the Elementary Effects method. The results show that the spatial Sobol' approach gives the best insight on AquiferSim behavior. OAT local variations of inputs around some sample points allow checking of the robustness of model predictions around those points, but give no insight on the relative importance of inputs. The Elementary Effects method shows that land use layer is the most influential input factor, but fails to capture interactions between input factors. The spatial Sobol' approach identifies the land use layer as being the most influential. It shows that strong interactions occur between most of the inputs, explaining 43% of the output variability.
Keywords: sensitivity analysis, Sobol, Elementary Effects