Breaking out from a restricted range: Alternative habitat models to assess population perspectives

Abstract

Habitat conservation for restricted-range species should also consider adjacent areas, but the analytical approaches for such assessments (particularly for a future perspective) are constrained by currently observed habitat relationships. We used two conceptually different habitat modelling approaches for analysing habitat distribution for the isolated Estonian population of a species of European conservation concern, the Siberian flying squirrel (Pteromys volans (Linnaeus, 1758)). We expected that the correlative (statistical) approaches based on current location data will increasingly deviate along with the distance from the current range, compared with a mechanistic approach based on limiting factors for the species. For conservation planning, we also investigated how the current protected area network covers quality habitats around the current range. We constructed three alternative correlative models (MaxEnt; Random forest; Generalized Boosted Regression) utilizing remote-sensing (Sentinel-2; LiDAR) and forest inventory data for 1299 occurrences in the currently occupied ca. 1400 km2 range. A mechanistic model was constructed as a decision tree that distinguished 11 quality classes of forest land based on the ecological prioritization of limiting factors: site type; forest cover; abundance of key tree species; stand age; patch size; and layer structure. Supporting our expectation, an overall good accordance of habitat predictions of all the correlative models and the mechanistic model (at 30 × 30 m pixel size) declined with the distance from the current range. The MaxEnt model most closely followed the full range of habitat quality classes of the mechanistic model, while the other correlative models emphasized the highest habitat-quality class. Within the current range, both MaxEnt and the mechanistic model similarly revealed habitat quality differences between occupied and unoccupied species protection areas. Delineation of habitat aggregations all over the country based on the mechanistic model revealed habitat loss both within and adjacent to the current range, which sets limits to local population recovery. For analysing wider options, we recommend complementing statistical spatial modelling of current conditions with ecologically sound mechanistic approaches. Based on our specific case, we outline how such model predictions can be assessed for management planning beyond current range.