Brice B. Hanberry
Spatial thinning of samples to remove clustered georeferenced records at local scales has become standard practice for species distribution models. Alternatively, I demonstrated the sampling issue of unequal sampling at global scales, and provided and tested a solution by equalizing samples. First, I used 17,200 species to characterize intensity gradients, with kernel density grids of georeferenced records, for 12 taxa. Secondly, I developed a method to even sampling intensity, by dividing kernel density grids into 10 quantiles of sampling intensity classes by taxa, and for each species, I determined the midpoint value of sample number in sampling intensity classes and removed records that exceeded the midpoint number. Thirdly, I compared species distribution models from equalized, unthinned, and thinned records, including locational distance uncertainty of records, to models from range maps. Only bird species deviated from greatest sampling intensity in Europe. Equalization reduced 42% of 140 million records to 15% of 50 million records in the greatest intensity sampling class. Models from equalized sampling maximized both the intersection and area predicted as present with models from range maps, with limited overprediction relative to models from thinned records. No sample processing approach generated models with the coverage of wide-ranging species. Models from any distance uncertainties were similar, but inclusion of records with distance uncertainties of 5 km resulted in even greater concentration of records in Europe and predicted area concentrated in Europe, for non-equalized records. Models from equalized sampling were able to avoid greater concentration of samples and omission error, due to recruiting records from a range of distance uncertainties in less sampled areas to increase predicted areas of presence, relative to unthinned samples, and reduce commission error, relative to thinned samples.