Model limitations, and methodological problems
Recently, numerous SDM have been based on presence-only data and employ so-called background points (pseudo-absences). Nonetheless, data not only on species presence but also their true (i.e., confirmed) absence are considered more relevant for modelling (Brotons et al., 2004; Barbet-Massin et al., 2012; Elith et al., 2020). Unfortunately, confirmed absence data are problematic because they need a high sampling effort (Barbet-Massin et al., 2012; MacKenzie & Royle, 2005) to be realistic. Our results show that in spite of the high-quality data employed here, exclusion of squares with a richness of neophytes (considered here as target species group) improves the model’s performance. This suggests an issue of sampling bias, which can be mediated by appropriate procedures. Our approach seems to be promising, but it needs further study in order to better understand its operation. The typical assumption, such as higher sampling effort in densely populated areas and near roads, is not adequate for invasive species because they typically occur in urban areas and along communication routes (Niinemets & Peñuelas 2008; Szymura et al., 2016).
Another problem consists of causality in our model: the approach applied represents a correlative type of model that is unable to directly capture the underlying processes driving the observed patterns of distribution. Contrary to this, the mechanistic (or process-based) models, which are built using explicit descriptions of biological mechanisms, are free from this disadvantage (Yates et al., 2018). However, they need appropriate formulation including detailed data on species response to environment, preferably coming from experiments, which are typically unavailable. In practice, the models rely to a considerable degree on parametrization based on observational data, and as a result, the difference between correlative and mechanistic models is often fuzzy (Yates et al., 2018). To conclude, regarding the recent state of knowledge regarding processes driving Solidago invasion, the mechanistic models do not have a lot of advantages compared with correlative models, especially given the lack of data for parametrization. Such data will come from experiments, not from observational study.