Finding climatically analog but independent locations for model evaluation
Climate is an important driver of ecological processes and finds widespread application in species distribution models (SDM). Deciding whether a model applies to a location requires knowing whether the location is climatically analog to the locations used for model building and calibration, because extrapolating renders a model invalid. Such knowledge is useful for determining whether a) a location can be used for model evaluation; b) a location is within the native range climatic niche of an invasive species; or c) non-analog communities assembled under non-analog climate. I compared the performance of four different approaches to determining whether locations are climatically analog: 1) multidimensional cube; 2) multidimensional minimum convex hull; 3) mahalanobis distance; and 4) a combination of PCA and cluster analysis. Creating and calibrating SDM’s is often a complicated multi-step process and the only valid evaluation is to test predictions on independent locations. Therefore, I compared the four methods by their effect on the evaluation of distribution models.
The multidimensional convex hull performed best at discriminating interspersed test locations that should have similar climate to training locations, from spatially segregated test locations. However, all methods performed adequately when adjusted well. The cube was least discriminant and the PCA and cluster analysis combination was by far the most complicated and sensitive to adjustments. In each case, climate variable selection used for discrimination was important and tolerances could be introduced, greatly helping the fine tuning. In particular the minimum convex hull and the mahalanobis distance were sensitive to inclusion of large numbers of variables, suggesting that finding locations to be non-analog is a strong function of the number of dimensions considered. There is no objectively right approach. Researchers must pick their technique based on intended application and have much scope for fine-tuning it to their need. Therefore, it is important to have a system for defining the need and testing whether the methods fulfill the intended need before picking a method. The principles demonstrated herein are equally applicable to other multidimensional important factors.