Combining habitat and spatial data improves predictions of Miscanthus occurrence

Background/Question/Methods

The introduction of novel crops for bioenergy production has sparked concerns over the risk of fostering new plant invasions. Tying data on species biology to landscape attributes is crucial for predicting invasion risks and optimizing management strategies across a range of potentially vulnerable habitats. However, basic information on species thresholds and limitation may not directly translate into large scale predictions about invasion spread. Unclear relationships between drivers at different spatial scales may obscure links between species invasiveness and spatially explicit habitat factors. Further, dispersal potential also influences the relative risk across a landscape, consequently highlighting uncertainties in distinguishing likely colonization sites from suitable habitat. We used transect survey data to evaluate the strength of logistic models based on small scale habitat studies to predict occurrence of Miscanthus sinensis, a candidate bioenergy crop, across two southeastern landscapes.  We measured habitat factors previously associated with M. sinensis performance that could also be linked to large scale remotely sensed variables, and recorded presence and absence of M. sinensis, to test predictions of M. sinensis occurrence. We further examined how the addition of a spatial variable based on an empirically-derived dispersal kernel influenced the accuracy of model predictions.

Results/Conclusions

Light availability, as represented by canopy openness, was the best habitat predictor of M. sinensis, and substituting openness derived from remote sensing for field measurements of overhead canopy cover reduced predictive capacity by less than 10%. We were consistently able to accurately predict greater than 80% of M. sinensis absences; however, presence predictions using canopy openness retained around 30% accuracy. Adding the spatial variable more than doubled the success of presence predictions from the initial 30% to over 70%, while consistently correctly identifying M. sinensis absences. Therefore, local scale habitat measurements can be linked to data at a larger scale to generate predictions of M. sinensis occurrence, but adding spatial information substantially increased prediction success. Additionally, though our capacity to predict M. sinensis presence depended heavily on information associated with known M. sinensis sites, models were consistently strong in identifying areas where M. sinensis was absent.  Consequently, focusing on the distribution of unsuitable, rather than highly vulnerable, habitats in calculations of invasion potential may improve risk predictions for Miscanthus cultivation.