LUISA (Library User Interface for Sensitivity Analysis): a generic software component for sensitivity analysis of bio-physical models

Sensitivity analysis is crucial for a better understanding of model behavior, both for developers and users. Developers can be supported in avoiding over parameterizations and in focusing their attention only on the processes with a significant impact on the output(s) of interest. Model users can benefit from sensitivity analysis by identifying the most relevant parameters in a specific either biophysical or physical context and, therefore, in optimizing the available resources for determining their values, by direct measurements or calibration. When biophysical, deterministic models are run with stochastic data like weather of climate ensembles, and when other inputs, such as management actions, of a model vary substantially, the results of a sensitivity analysis may provide information about different, site specific strategies for operational use. The availability of a generic software component integrated into the modeling and simulation software environment would hence allow the estimation of differences in the behavior of models in different contexts. It is possible to classify the methods for sensitivity analysis developed in the last decades in three groups: the one-factor-at-a-time method, the methods based on regression and the variance-based Monte-Carlo methods. The first group is represented as an example by Morris' method, which calculates two metrics: the average (mu) and the standard deviation (s) of the population of the incremental ratios according to an opportune generation of a sample of the possible combination of parameters. The most famous methods belonging to the second group are the Latin Hypercube, the Random and the Quasi-random Lp-Tau. They differ in the method used for generating the sample, while are all based on a linear regression between the differences in the outputs of the model and those in the values of parameters to calculate sensitivity indices. The third group is based on the decomposition of the total variance in terms of increasing dimensionality and it is able to quantify the effect of the interactions among parameters. The methods based on this principle are Fourier Amplitude Sensitivity Test (FAST), Extended FAST, and Sobol's method. The last group, and in particular Sobol's method, is considered the most powerful and precise in identifying the sensitivity of the model output. in response to changes in model parameters. Their drawback is the computational cost since they involve the estimation of k-dimensional integrals. On the other hand, the first group methods are the one requiring the smallest sample for ranking the parameters according to their relevance and it is considered particularly suitable for preliminary screenings of models with several parameters. This paper describes the LUISA (Library User Interface for Sensitivity Analysis) component, based on the SimLab (http://simlab.jrc.ec.europa.eu/) C++ DLL. LUISA has been developed in C# under the. NET platform, with the goal to facilitate the integration of sensitivity analysis capabilities into bio-physical model frameworks. To provide illustrative case studies, a spatially distributed sensitivity analysis of two different biophysical models was carried out using the MARS database (http://mars.jrc.ec.europa.eu/), covering the pedo-climatic conditions of Europe. The two models used were the WARM model for rice simulations and the generic crop simulator CropSyst with a parameterization for wheat. Results are presented and discussed according to the spatial variability of parameter relevance.