With spatially distributed models, the effect of spatial uncertainty of the model inputs is one of the least understood contributors to output uncertainty and can be a substantial source of errors that propagate through the model. The application of the global uncertainty and sensitivity (GUA/SA) methods for formal evaluation of models is still uncommon in spite of its importance. Even for the infrequent cases where the GUA/SA is performed for evaluation of a model application, the spatial uncertainty of model inputs is disregarded due to lack of appropriate tools. The main objective of this work is to evaluate the effect of spatial uncertainty of model inputs on the uncertainty of spatially distributed watershed models in the context of other input uncertainty sources. A new GUA/SA framework is proposed in this dissertation in order to incorporate the effect of spatially distributed numerical and categorical model inputs into the global uncertainty and sensitivity analysis (GUA/SA). The proposed framework combines the global, variance-based method of Sobol and geostatistical techniques of sequential simulation (SS). Sequential Gaussian simulation (SGS) is used for estimation of spatial uncertainty for numerical inputs (like land elevation), while sequential indicator simulation (SIS) is used for assessment of spatial uncertainty of categorical inputs (like land cover type). The Regional Simulation Model (RSM) and its application to WCA-2A in the South Florida Everglades is used as a test bed of the framework developed in this dissertation. The RSM outputs chosen as metrics for GUA/SA for this study are key performance measures generally adopted in the Everglades restoration studies: hydroperiod, water depth amplitude, mean, minimum and maximum. The GUA/SA results for two types of outputs, domain-based (spatially averaged over domain) and benchmark cell-based, are compared. The benchmark cell-based outputs are characterized with larger uncertainty than their domain-based counterparts. The uncertainty of benchmark cell-based outputs is mainly controlled by land elevation uncertainty, while uncertainty of domain-based outputs it also attributed to factors like conveyance parameters. The results indicate that spatial uncertainty of model inputs is indeed an important source of model uncertainty.

The land cover distribution affects model outputs through delineation of Manning’s roughness zones and evapotranspiration factors associated to the different vegetation classes. This study shows that in this application the spatial representation of land cover has much smaller influence on model uncertainty when compared to other sources of uncertainty like spatial representation of land elevation.

The spatial uncertainty of land cover was found to affect RSM domain-based model outputs through delineation of Manning’s roughness zones more than through ET parameters effects.

The relationship between model uncertainty and alternative spatial data resolutions was studied to provide an illustration of how the procedure may be applied for more informed decisions regarding planning of data collection campaigns. The results corroborate a proposed hypothetical nonlinear, negative relationship between model uncertainty and source data density. The inflection point in the curve, representing the optimal data requirements for the application, is identified for the data density between 1/4 and 1/8 of original data density. It is postulated that the inflection point is related to the characteristics of the spatial dataset (variogram) and the aggregation technique (model grid size).

The framework proposed in this dissertation could be applied to any spatially distributed model and input, as it is independent from model assumptions.