Résumé
Variance-based Sobol' global sensitivity analysis (GSA) was initially designed for the study of models with scalar inputs and outputs, while many models in the environmental field are spatially explicit. As a result, GSA is not a common practise in environmental modelling. In this paper we describe a detailed case study where GSA is performed on a spatially dependent model for flood risk economic assessment on the Orb valley (southeast France). Spatial input factors are handled by associating randomly generated map realizations to scalar values sampled from discrete uniform distributions. The realisations of random input maps can be generated by any method including geostatistical simulation techniques, allowing for spatial structure and auto-correlation to be taken into account. The estimation of sensitivity indices on ACB-DE spatial outputs makes it possible to produce maps of sensitivity indices. These maps describe the spatial variability of Sobol' indices. Sensitivity maps of different resolutions are then compared to discuss the relative influence of uncertain input factors at different scales.