COS 197-3 - Validating graph-based connectivity models for a forest tropical bird species using independent presence/absence and genetic datasets

Background/Question/Methods

Habitat connectivity is commonly modeled by landscape graphs, i.e. sets of habitat patches (nodes) connected by potential dispersal paths (links). Graph-based distances and connectivity metrics are then integrated into conservation or knowledge-driven approaches. Because expert opinion or species distribution models (SDM) are most often used for delineating patches and assigning every landscape feature a resistance to movement, graphs lack of validation from field data closely reflecting functional connectivity. Accordingly, we aimed to answer the following question: are landscape graphs reflecting ecological processes influenced by habitat connectivity? To that purpose, we modeled the habitat network of a forest bird (Setophaga plumbea) in the Guadeloupe island with graphs built from either expert opinion, or from specialization indices or a SDM computed with 991 presence/absence observations. In parallel, we genotyped 712 birds from 27 populations (12 microsatellites). We then computed genetic indices (allelic richness, population-level genetic differentiation) and pairwise genetic distances (FST). We finally studied the relationships between (i) genetic distances or indices and (ii) cost-distances or connectivity metrics, using MLPE distance models and Spearman correlations between metrics. We paid attention to the sensitivity of the models to the topology of the considered population pairs, and validated them with a cross-validation approach.

Results/Conclusions

Overall, the landscape graphs reliably reflected the influence of connectivity on population genetic structure. The validation R² of distance models reached values up to 0.30, with substantial variations according to the topology of the population pairs included in the model. Besides, correlation coefficients between connectivity metrics and genetic indices reached values up to 0.71 and were less variable. Yet, the relationship between graph ecological relevance, data-requirements and construction and analysis method complexity was not straightforward. The landscape graph based upon the most complex construction method (SDM) had sometimes a lower ecological relevance than the others. Cross-validation methods and sensitivity analyses allowed us to make the advantage and limitation of every construction method spatially-explicit. In sum, we confirmed the relevance of landscape graphs for conservation modeling but we call for a case-specific consideration of the cost effectiveness of their construction methods.