Extending coverage and thematic resolution of compositional land cover maps in a hierarchical Bayesian framework

Ecological models are constrained by the availability of high-quality data at biologically appropriate resolutions and extents. Modeling a species' affinity or aversion with a particular land cover class requires data detailing that class across the full study area. Data sets with detailed legends (i.e., high thematic resolution) and/or high accuracy often sacrifice geographic extent, while large-area data sets often compromise on the number of classes and local accuracy. Consequently, ecologists must often restrict their study extent to match that of the more precise data set, or ignore potentially key land cover associations to study a larger area. We introduce a hierarchical Bayesian model to capitalize on the thematic resolution and accuracy of a regional land cover data set, and on the geographic breadth of a large area land cover data set. For the full extent (i.e., beyond the regional data set), the model predicts systematic discrepancies of the large-area data set with the regional data set, and divides an aggregated class into two more specific classes detailed by the regional data set. We illustrate the application of our model for mapping eastern white pine (Pinus strobus) forests, an important timber species that also provides habitat for an invasive shrub in the northeastern United States. We use the National Land Cover Database (NLCD), which covers the full study area but includes only generalized forest classes, and the NH GRANIT land cover data set, which maps White Pine Forest and has high accuracy, but only exists within New Hampshire. We evaluate the model at coarse (20 km² ) and fine (2 km² ) resolutions, with and without spatial random effects. The hierarchical model produced improved maps of compositional land cover for the full extent, reducing inaccuracy relative to NLCD while partitioning a White Pine Forest class out of the Evergreen Forest class. Accuracy was higher with spatial random effects and at the coarse resolution. All models improved upon simply partitioning Evergreen Forest in NLCD based on the predicted distribution of white pine. This flexible statistical method helps ecologists leverage localized mapping efforts to expand models of species distributions, population dynamics, and management strategies beyond the political boundaries that frequently delineate land cover data sets.