Sensitivity of a white-tailed deer habitat-suitability index model to error in satellite land-cover data: implications for wildlife habitat-suitability studies

Abstract

Biologists are increasingly using classified satellite data in habitat analyses, especially due to the recent availability of the National Land Cover Dataset (NLCD) and Gap Analysis Project (GAP) classified maps. However, this type of land-cover data may contain substantial class error incorporated during the classification process. Accuracy assessments provide a measure of class-specific error, but they do not provide information on the spatial distribution of error. Error simulation procedures allow users to gauge the sensitivity of their analyses output to errors in the input land-cover data, although these procedures tend to be computationally intensive. We used pixel- and patch-based error simulation to evaluate the effect of class error in the NLCD on the calculation of habitat-suitability index (HSI) values for white-tailed deer (Odocoileus virginianus) in a high-suitability (agricultural) landscape and a low-suitability (forested) landscape. We incorporated error into landscapes at rates reported in the NLCD accuracy assessment for New York and New Jersey using 2 techniques: pixel-based simulations, in which individual pixels were changed from one cover type to another, and patch-based simulations, in which entire patches were changed. Resulting HSI values were higher in all simulated landscapes than in the original landscapes. The largest increase in HSI, 0.57, occurred using a pixel-based simulation in the low-suitability landscape, due partially to increased interspersion and creation of new one-pixel patches. Patch-based simulations provided a more conservative estimate of the effect of error, an increase in HSI of 0.07–0.28. The difference in HSI between pixel- and patch-based error simulations suggests that the spatial distribution of error in land-cover data may strongly affect the calculation of HSI, especially if the model contains habitat variables related to landscape configuration. We suggest that biologists who make use of classified land-cover data evaluate their analyses for potential sensitivity to error, examine confusion matrices provided with accompanying metadata, and conduct error simulations when feasible.