Comparison of landscape graph modelling methods for analysing pond network connectivity

Context Landscape fragmentation negatively impacts species populations by isolating them. Assessing landscape connectivity could help to improve biodiversity conservation. Among various methods available to model and analyse connectivity, graph-theoretic approaches are recognized as powerful tools, even if their ecological significance may be questionable in some cases. Indeed, there are many ways to construct a landscape graph and their impacts on the assessment of connectivity are rarely explored. Objectives Our aim was to compare three methods of constructing landscape graphs to identify differences and similarities in the resulting network connectivity. The methods can be distinguished according to the type of data used: expert opinions, field data or a combination of the two. The methodological framework was applied to seven pond-dwelling species ( Alytes obstetricans, Bufo bufo, Epidalea calamita, Hyla arborea, Natrix natrix, Rana temporaria, Triturus cristatus ) in the Ile-de-France region (France). Methods Three common methods were applied to construct landscape graphs: (1) using a land cover map (LM) and expert opinions to define nodes and links; (2) using a habitat suitability model (HSM) and species occurrence data to define nodes and links; and (3) using a HSM to define nodes and a land cover map to define links (HSM_LM). To carry out our study, we produced a land cover map, collected and prepared input data for HSMs, generated HSMs to map the probability of species occurrence and constructed landscape graphs from the three methods. For each of them, several connectivity metrics were calculated and compared. Results The results revealed large differences in the statistical distribution of connectivity values, even though the spatial location of the main areas of low and high connectivity was roughly the same. In general, the LM method provided lower values of connectivity and smaller areas of high values than the other two, regardless of species. Conversely, the HSM method had the highest connectivity values, while the combined HSM_LM method appeared to be intermediate. Conclusions Our study was not intended to conclude whether one method is better than another; only to point out that results vary greatly depending on the graph construction method. To evaluate the predictive performance of each model, a validation process should be conducted with another independent biological dataset, which was not available in our study. The high variability