Sensitivity of habitat network models to changes in maximum dispersal distance.

Predicting the presence or absence (occurrence-state) of species in a certain area is highly important for conservation. Occurrence-state can be assessed by network models that take suitable habitat patches as nodes, connected by potential dispersal of species. To determine connections, a connectivity threshold is set at the species' maximum dispersal distance. However, this requires field observations prone to underestimation, so for most animal species there are no trustable maximum dispersal distance estimations. This limits the development of accurate network models to predict species occurrence-state. In this study, we performed a sensitivity analysis of the performance of network models to different settings of maximum dispersal distance. Our approach, applied on six amphibian species in Switzerland, used habitat suitability modelling to define habitat patches, which were linked within a dispersal distance threshold to form habitat networks. We used network topological measures, patch suitability, and patch size to explain species occurrence-state in habitat patches through boosted regression trees. These modelling steps were repeated on each species for different maximum dispersal distances, including a species-specific value from literature. We evaluated mainly the predictive performance and predictor importance among the network models. We found that predictive performance had a positive relation with the distance threshold, and that almost none of the species-specific values from literature yielded the best performance across tested thresholds. With increasing dispersal distance, the importance of the habitat-quality-related variable decreased, whereas that of the topology-related predictors increased. We conclude that the sensitivity of these models to the dispersal distance parameter stems from the very different topologies formed with different movement assumptions. Most reported maximum dispersal distances are underestimated, presumably due to leptokurtic dispersal distribution. Our results imply that caution should be taken when selecting a dispersal distance threshold, considering higher values than those derived from field reports, to account for long-distance dispersers.

Predicting species occurrences with habitat network models.

Biodiversity conservation requires modeling tools capable of predicting the presence or absence (i.e., occurrence-state) of species in habitat patches. Local habitat characteristics of a patch (lh), the cost of traversing the landscape matrix between patches (weighted connectivity [wc]), and the position of the patch in the habitat network topology (nt) all influence occurrence-state. Existing models are data demanding or consider only local habitat characteristics. We address these shortcomings and present a network-based modeling approach, which aims to predict species occurrence-state in habitat patches using readily available presence-only records.For the tree frog Hyla arborea in the Swiss Plateau, we delineated habitat network nodes from an ensemble habitat suitability model and used different cost surfaces to generate the edges of three networks: one limited only by dispersal distance (Uniform), another incorporating traffic, and a third based on inverse habitat suitability. For each network, we calculated explanatory variables representing the three categories (lh, wc, and nt). The response variable, occurrence-state, was parametrized by a sampling intensity procedure assessing observations of comparable species over a threshold of patch visits. The explanatory variables from the three networks and an additional non-topological model were related to the response variable with boosted regression trees.The habitat network models had a similar fit; they all outperformed the non-topological model. Habitat suitability index (lh) was the most important predictor in all networks, followed by third-order neighborhood (nt). Patch size (lh) was unimportant in all three networks.We found that topological variables of habitat networks are relevant for the prediction of species occurrence-state, a step-forward from models considering only local habitat characteristics. For any habitat patch, occurrence-state is most prominently influenced by its habitat suitability and then by the number of patches in a wide neighborhood. Our approach is generic and can be applied to multiple species in different habitats.