In situ conservation – the protection of species in their natural habitats – is the most powerful biodiversity conservation strategy, but protected areas cannot be expected to conserve what is not represented in them in the first place. Historically, most protected areas were created opportunistically on an individual basis (Pressey & Tully, 1994), but the networks obtained from this piecemeal approach are often incomplete in their representation of biodiversity (e.g. Scott et al., 2001; Rodrigues et al., 2004b), and inefficient in their use of land available to conservation (e.g., Pressey, 1994; Fuller et al., 2010). Systematic conservation planning emerged from the recognition that, given limited conservation resources, protected areas need to be selected not just based on their individual characteristics but as coherent networks of complementary sites (Pressey et al., 1993; Margules & Pressey, 2000). Yet this approach is extremely data-hungry, requiring data on the spatial distribution of all biodiversity features of interest, across all candidate sites (Margules & Pressey, 2000), which seriously limits its applicability to the regions of the world that need it the most: biodiversity-rich, data-poor regions, often with the least developed protected area networks (Pimm, 2000).
Species distribution data in such regions are typically available in one of two forms – extent of occurrence (EOO) data, or point locality records (Gaston, 1994) – that suffer from contrasting limitations in their value to conservation planning (Rondinini et al., 2006). EOO data are generalized polygons of plausible range, often obtained through interpolation from point records, and may include relatively extensive areas from which the species is absent (e.g. freshwater species mapped as continuous EOO polygons covering both freshwater terrestrial habitats). EOO data typically overestimate (sometimes vastly so) the area occupied by each species (Jetz, Sekercioglu & Watson, 2008), resulting in high levels of commission errors (false presences) whereby species are assumed to be protected in sites where in fact they do not occur (Rondinini et al., 2006). Point locality data are obtained from recent records of confirmed species presence, often very incomplete and biased representations of species' true area of occupancy. When applied to conservation planning, they result in high levels of omission errors (false absences), whereby species are assumed absent from places they actually occur (Rondinini et al., 2006). For conservation purposes, it is more important to minimise commission errors than omission errors (Rodrigues et al., 2004b), because assuming a species to be conserved when it is not may ultimately result in its loss. However, omission errors affect the efficiency of systematic conservation planning, by reducing the spatial options available to the planer (Rondinini et al., 2006).
In this issue, Beresford and colleagues propose a method for obtaining a compromise between these two types of data: they refined EOO data of 157 globally threatened bird species (BirdLife International, 2008) into maps of extent of suitable habitat (ESH). Whereas much more sophisticated models of species distribution can be obtained using other methods (e.g. Segurado & Araujo, 2004), this approach is appealing for practical conservation planning because it combines data – coarse EOO polygons, descriptive habitat information and satellite land cover data – that are readily available for many species in data-poor regions (Bartholomé & Belward, 2005; IUCN, 2010). Beresford et al. (2011) were conservative, including all habitat types associated with each species, yet they found species' range was reduced in average to 28% of its original EOO.
The most interesting result in Beresford et al. (2011) comes, in my opinion, from their use of an independent dataset to ground-truth the results of EOO and ESH species maps: records of threatened species in important bird areas (IBAs) of Africa (Fishpool & Evans, 2001). IBAs are a network of key sites for the conservation of birds, selected based on the presence of globally threatened, rare and migratory or congregatory species. Given that all IBAs are specifically scrutinized for the presence of threatened species, they are a reliable dataset not only on the presence but also on the absence of each species in each site. Beresford et al. (2011) then investigated the degree of agreement between the presence/absence of each species in each IBA as predicted from EOO and ESH maps, and the real occurrence according to the IBA data.
They found that EOO data significantly overestimate the presence of species in IBAs (table 1 in Beresford et al. 2011). Although high levels of commission errors were expected, it is nonetheless sobering that out of 3577 predicted presences only 847 (24%) are real according to observed IBA records. This reinforces the need for great caution in applying the results of conservation planning based on coarse EOO data: the regions highlighted should be priorities for finer-scale assessments based on much refined field data, not prescriptive results on the location of specific protected areas (Rodrigues et al., 2004a). More surprising is the result that the EOO data also suffer from important levels of omission errors: out of the 1003 observed presences in IBAs, the EOO data missed 156 (16%). This suggests considerable scope for improvement in the delimitation of the EOO boundaries, by making better use of available information.
The ESH approach was expected to reduce the commission errors found in the EOO data, by eliminating parts of species' ranges where habitat is clearly unsuitable. And indeed, the number of false presences was reduced considerably from 2730 with the EOO data to 1887 with the ESH data (table 1 in Beresford et al. 2011). Unfortunately, though, agreement between predicted and observed presences also declined, from 847 to 630 records, with a concomitant increase in the number of false absences (omission errors) to 373. As a result, 37% of all observed presences as per the IBA records are missed by the ESH, twice as many as in the EOO data. Disappointingly, it appears that the contraction in species' mapped ranges as EOO polygons were refined into ESH resulted in the removal of not only parts of the ranges where species are truly absent (as desirable, to reduce commission errors), but also parts of the range where the species is truly present (increasing omission errors). Overall, the fraction of presences predicted by the ESH confirmed by the IBA data (25%) is only marginally higher than the corresponding value with the EOO data (24%). This suggests very little discriminatory capacity of the ESH method to distinguish between occupied and non-occupied parts of species' EOO. Future analyses should test if it is significantly better than a simple random cropping of species' EOOs.
Further analyses are needed to clarify the reasons for such counter-intuitive result. For example, it is possible that the ESH approach performs badly for species from habitat types inadequately represented in land cover maps (e.g. freshwater species), but may be appropriate for others. In the meantime, the present study does not support the assumption that the ESH approach improves the quality of coarse species distribution data for application to regional conservation planning in relation to EOO maps. Overall, these results are a sober reminder of the limitations of large-scale species' distribution data in biodiversity-rich, data-poor, regions, even for the best-studied taxa.