Name: Silveira L

Biodeserts supervisors: Brito JC

Co-supervisors: Hugo Rebelo (FCUP)

Title: Validating the transferability of ecological models under global change scenarios with Holocene rock-art

Institution: University of Porto, Portugal

Status: Completed



The world is facing a global biodiversity crisis, with an increasing rate of species extinction due to habitat loss and climate change. Predictions of biodiversity responses to global change play a key role to guide conservation efforts. Predictions of climate change effects on biodiversity distribution usually rely on ecological niche-based models due to their ability to predict species responses in unsampled spatial and/or temporal scenarios. Assessing the quality of model transferability across time is critical if they are to be used to predict responses beyond the conditions under which they were trained. However, validating model transferability requires independent data that is impossible to obtain in past or future time scenarios. Rock-art depicting species may provide evidence on paleo-distributions and a unique opportunity to test the robustness of model projections, which may help understating the reliability of future projections. Using rock-art from the Holocene of African savannah elephant (Loxodonta africana) and Giraffe (Giraffa camelopardalis), this work evaluated the predictive ability of ecological models based in distinct: i) strategies to select pseudo-absence data (minimum distance to observations, random data, and environmentally restricted data); ii) sources of climatic layers (Worldclim and Paleoview) for three time periods (Current, 6,000BP and 7,900BP); and iii) environmental variables (land-cover and bioclimatic datasets). Models were constructed with algorithm Random Forests for the three time periods (Current, 6,000BP and 7,900BP) and projected to all possible combinations of time periods to predict suitable areas for both species. Overall, models calibrated with environmentally restricted pseudo-absences performed better, no major differences were found between models calibrated with different sources of bioclimatic layers, and models coupling land-cover variables with bioclimatic variables performed better than pure bioclimatic or land-cover based models. The study here developed allows providing a series of recommendations related to model transferability in time: 1) When the distributions of studied species are not totally known but the ecological niche traits are well known, it is recommend using environmentally restricted pseudo-absences datasets; 2) When projecting models, it is best to use Worldclim bioclimatic variables given the higher availability of spatial resolutions and GCMs; 3) Coupling land-cover together with bioclimatic variables is recommended, since the association of these variables may reflect accurately environmental changes between calibration and projection periods. Still, given the limitations in past land-cover availability, trade-offs will need to be established between pixel sizes and data sources.