A single-algorithm ensemble approach to estimating suitability and uncertainty: cross-time projections for four Malagasy tenrecs

Aim: Ecological niche models (ENMs) are used widely in ecology, evolution, global change biology, but model uncertainty remains an underappreciated issue. Generally, either a single model from one algorithm or an ensemble of single models from different algorithms is used to provide a prediction. In addition to variability among algorithms, recent studies have shown the need to consider variability within a single algorithm, for example optimizing model complexity by tuning model settings. We present an ensemble ENM using a single-algorithm approach, while adjusting model settings to maximize performance. Location: Madagascar. Methods: We used maxent, bioclimatic variables and occurrence records of four species of Malagasy tenrecs (Family Tenrecidae). We calibrated and evaluated preliminary models using a jackknife approach, tuning two model settings to estimate optimal model complexity. We chose a suite of top-performing preliminary models and then generated a consensus prediction. Furthermore, we calculated the variability among predictions of the co-optimal models to indicate variation in geography (i.e. uncertainty). We then did the same after projecting the predictions to climatic estimates for the Last Glacial Maximum and the year 2070. Results: The default settings were never identified as optimal for any of the four species. The model settings considered as the co-optimal solutions essentially led to the same evaluation statistics; however, they showed high variation in their geographic predictions for three of the four species. Additionally, variation among such models was greater when transferred across time. Main conclusions: This approach likely can provide better predictions for a single algorithm as well as quantifications of within-algorithm uncertainty, qualities that are highly useful in interpreting reconstructed suitable areas or forecasts of potential range shifts under future climate change. Finally, this within-algorithm uncertainty can be integrated into a larger framework that considers variability due to other factors (e.g. related to input data, alternate algorithms or various Global circulation models).