Overall, the predictions made here reflect previous findings about potential A. macrophyllum habitat changes (Hamdan & Schmidt, 2012; Hamann & Wang, 2006), and the model accurately reflects what we know about its current range and habitat needs (Fryer, 2011). However, there are some limitations introduced by data availability and the modelling method.
The training set used pseudo-absence points to reflect unsuitable habitat, but observed absence training data would be more representative of actual habitat suitability. Unfortunately this type of data is not typically available or collected outside of specific applications, and absence data collection is outside the scope of this course. It is also possible that the bioclimatic and topographic variables considered do not accurately reflect all controls on habitat suitability, as factors related to soil properties, land cover, and hydrology were not explicitly included.
As the random forest modelling framework is relatively new in ecological modelling, there are some uncertainties about its ability to model species distributions, and particularly its transferability. Heikinnen, Marmion & Luoto (2012) found that despite its high interpolation accuracy, random forest does not perform as well as other methods in extrapolating a model to a new set of input parameters, as is done in climate change response models. This contradicts Mi et al. (2017), who found random forest to be the most transferable of a variety of modelling methods. The reason for discrepancies in transferability between different applications and datasets is unclear. Further work is necessary to evaluate the merit of using random forest for predicting habitat changes, and the conditions under which extrapolation is a valid approach.
It is also worth noting that plant physiology, forest disturbance and management strategies are important controls in the dispersal over a species’ suitable range. The modelled distributions cannot reflect actual dispersal patterns, but rather only indicate a possible extent in the absence of other controlling factors. Further work could assess these limitations and their potential impact on A. macrophyllum dispersal into interior forests as predicted here.