Using local maps of spatial accuracy for benthic habitat models. GeoComputation 2019

Spatial predictive models are used extensively in ecological applications as well as in many other domains. Their effectiveness for a given purpose depends on, among other things, a robust assessment of accuracy, particularly classification accuracy, which quantifies the difference between a prediction/classification and its “true” or observed value. In spite of their limitations, overall accuracy metrics such as percent correctly classified and kappa are still widely used, although more precise metrics that are able to differentiate between errors associated with location versus quantity have been introduced (Pontius and Millones 2011, Pontius and Santacruz 2014). In addition to overall accuracy measures, different aspects of classification accuracy can be measured relative to the observed values (e.g., errors of omission, producer’s accuracy), the predicted values (e.g., errors of commission, user’s accuracy), and class-specific accuracy. While a variety of accuracy metrics are available, most provide a single assessment value for an entire study area without taking into consideration local variations in spatial accuracy. Global measures of accuracy can be misleading when accuracy varies spatially and may be particularly problematic when the relative importance of predictive accuracy is different throughout the study area. Recent developments in accuracy assessment from remote sensing have illustrated the benefits of providing local measures of accuracy, but these have yet to be applied within the ecological domain. We use a case study of benthic habitat classes in a remote region offshore from NW Australia to demonstrate the value of calculating locally varying diagnostics of classification accuracy for an ecological model by comparing the variation in the resulting maps to the associated global values.