Aim: Systematic conservation planning is vital for allocating protected areas given the spatial distribution of conservation features, such as species. Due to incomplete species inventories, species distribution models (SDMs) are often used for predicting species' habitat suitability and species' probability of occurrence. Currently, SDMs mostly ignore spatial dependencies in species and predictor data. Here, we provide a comparative evaluation of how accounting for spatial dependencies, that is, autocorrelation, affects the delineation of optimized protected areas.Location: Southeast Australia, Southeast U.S. Continental Shelf, Danube River Basin.
Methods:We employ Bayesian spatially explicit and non-spatial SDMs for terrestrial, marine and freshwater species, using realm-specific planning unit shapes (grid, hexagon and subcatchment, respectively). We then apply the software gurobi to optimize conservation plans based on species targets derived from spatial and non-spatial SDMs (10%-50% each to analyse sensitivity), and compare the delineation of the plans.Results: Across realms and irrespective of the planning unit shape, spatially explicit SDMs (a) produce on average more accurate predictions in terms of AUC, TSS, sensitivity and specificity, along with a higher species detection probability. All spatial optimizations meet the species conservation targets. Spatial conservation plans that use predictions from spatially explicit SDMs (b) are spatially substantially different compared to those that use non-spatial SDM predictions, but (c) encompass a similar amount of planning units. The overlap in the selection of planning units is smallest for conservation plans based on the lowest targets and vice versa.Main conclusions: Species distribution models are core tools in conservation planning. Not surprisingly, accounting for the spatial characteristics in SDMs has drastic impacts on the delineation of optimized conservation plans. We therefore encourage practitioners to consider spatial dependencies in conservation features to improve the spatial representation of future protected areas.
K E Y W O R D SBayesian hierarchical modelling, connectivity, gurobi, integer linear programming, spatial autocorrelation, spatial unit | 759 DOMISCH et al.