Skip to Main content Skip to Navigation
Conference papers

Is there room to optimise the use of geostatistical simulations for sensitivity analysis of spatial models?

Abstract : Complex spatial models are developed to support decision making in various fields of environmental management. These models use environmental data that is spatially distributed, including 2D fields derived from sampled data. These spatial inputs are always partly uncertain. In order to provide confidence in these models, it is necessary to assess the impact of all input uncertainties. Performing sensitivity analysis (SA) can achieve this goal: SA techniques will allow the robustness of the model predictions to be checked and the input factors that account for most of the model output variability to be identied. The variance-based Sobol' approach is one of the few sensitivity analysis techniques that is suitable for complex models with spatially distributed inputs, non-linear effects and interactions between factors. It is based on a wide random exploration of the uncertainty space of the input factors,through Monte-Carlo sampling schemes. Spatially distributed inputs can be included in the analysis by associating randomly generated map realizations to scalar values. This allows complex description of spatial uncertainty to be applied. Spatial uncertainty on continuous fields, such as land elevation or rainfall elds, can come from diverse sources: measurement errors in data acquisition, interpolation errors, etc. To simulate this uncertainty, geostatistical unconditional or conditional simulations can be used to generate random realizations of the continuous 2D elds used as model inputs. Yet spatial Sobol' methods need a large number of model runs to compute sensitivity indices. With time consuming models, it is necessary to get the most accurate sensitivity indices with the fewest model runs. One way to reach better efficiency in the computation of sensitivity indices is by using a better sampling design in the space of the input factors. This issue has been widely discussed in the case of models with scalar inputs. Latin hypercube sampling (LHS) has been documented as an efficient option, along with other sampling schemes (quasi-random, winding stairs). Yet, in the case of sensitivity analysis of models with spatial inputs, the influence of the sampling scheme used to generate geostatistical simulations has not, up to now, been discussed. Some techniques have been developed to apply LHS to sample geostatistical simulations (e.g. Pebsema and Heuvelink, 1999 and Xu, 2005). Using a spatial sampling technique such as these may lead to better efficiency in the computation of sensitivity indices on spatial models. The aim of this paper is to compare the accuracy and bias of Sobol sensitivity indices on a model with continuous 2D spatial inputs, with a) different sampling designs of geostatistical simulations for spatial model inputs (simple random sampling, Latin hypercube sampling), b) increasing sample size.
Document type :
Conference papers
Complete list of metadata
Contributor : Migration Irstea Publications <>
Submitted on : Friday, May 15, 2020 - 5:15:51 PM
Last modification on : Friday, April 16, 2021 - 3:29:24 AM


  • HAL Id : hal-02593405, version 1
  • IRSTEA : PUB00029125


Nathalie Saint Geours, Jean-Stéphane Bailly, F. Grelot, C. Lavergne. Is there room to optimise the use of geostatistical simulations for sensitivity analysis of spatial models?. Accuracy2010, Jul 2010, Leicester, United Kingdom. pp.34. ⟨hal-02593405⟩



Record views