Metamodeling and global sensitivity analysis for computer models with correlated inputs: A practical approach tested with a 3D light interception computer model

Models of biophysical processes are often time-consuming and their inputs are frequently correlated. This situation of non-independence between the inputs is always a challenge in view of simultaneously achieving a global sensitivity analysis of the model output and a metamodeling of this output. In this paper, a novel practical method is proposed for reaching this two-fold goal. It is based on a truncated Polynomial Chaos Expansion of the output whose coefficients are estimated by Partial Least Squares Regression. The method is applied to a computer model for heterogeneous canopies in arable crops, aimed to predict crop:weed competition for light. We now have fast-running metamodels that simultaneously provide good approximations of the outputs of this computer model and a clear overview of its input influences thanks to new sensitivity indices.