Exploring the effects of land-use classification errors on the outputs of a coastal habitat model using global sensitivity and uncertainty analysis

Abstract

Watershed models which form the basis for implementing managerial schemes and decisions often use thematic maps as input data. The crucial role that these models serve makes it necessary to explore the uncertainty involved in using the thematic maps because they can affect the overall reliability of the models. Classification errors are one source of uncertainty when using these thematic maps which can be propagated into the output of the model. This study evaluated the effects of classification errors on the outputs of a habitat model called SLAMM applied to a portion of Santa Rosa Island in Eglin Air Force Base, Pensacola, FL. The technique presented in this study employed a two-step global sensitivity and uncertainty analysis in order to (1) determine the important input factors and processes that control SLAMM’s output uncertainty; and (2) quantify SLAMM’s global output uncertainty and apportion it to the direct contributions and interactions of the important factors. At higher classification errors, the variability in the output for swamp, inland fresh marsh, and salt marsh was predominantly driven by land cover classification as opposed to DEM vertical error for lower range elevation when classification error was only at 5%. As error increased, more input factors contributed to the variability of the output for low-lying habitats (salt marsh, tidal flat, and beach) and the barrier island itself. Interactions among input parameters were in general not present. Classification errors resulted in bimodality in the probability distribution function (pdfs) of swamp and inland fresh marsh, while multimodal distribution with three peaks was observed for salt marsh. As the classification error decreased, the pdfs of swamp and inland fresh marsh was transformed to a unimodal distribution, while that of the salt marsh to a bimodal distribution similar to the ones reported at 5% classification error. Different classification errors tend to produce different list of important input factors for some outputs. An accurate assessment of classification errors is therefore necessary to identify the correct important factors for research prioritization purposes. The technique presented in this study is model independent and can be adopted to evaluate the effects of errors in the input data on model outputs. Knowing the effects of such uncertainties on the output will result in a more reliable model and a sound managerial decision.