Effects of non-representative sampling design on multi-scale habitat models: flammulated owls in the Rocky Mountains

Sampling bias and autocorrelation can lead to erroneous estimates of habitat selection, model overfitting and elevated omission rates. We developed a multi-scale habitat suitability model of the flammulated owl (Psiloscops flammeolus) in the Northern Rocky Mountains based on extensive but spatially clustered survey data, and then used simulations to evaluate the effects of spatially non-representative and spatially representative sampling strategies on model performance and predictions. Our hypothesis was that models trained with spatially non-representative simulated datasets would suffer from bias in parameter estimates, and would show lower predictive performance. The models trained with the spatially representative simulated datasets greatly outperformed the models trained with the spatially non-representative simulated datasets judged on standard metrics of model performance. However, the spatially non-representative models produced superior predictions based on their ability to identify the correct spatial scales, covariates, signs and magnitudes of the species-environment relationships, when compared to the spatially representative models. Thus, it is likely that representative spatial sampling across a broad range of environmental gradients also resulted in over-dispersion of sampling data, with a higher proportion of samples falling in areas of low probability of presence, leading to lower ability to resolve the relationships between species presence-absence and environmental covariates. In contrast, the spatially non-representative sampling, by concentrating sampling along environmental gradients that are characterized by higher probability of presence of the modelled species, produced predictions that, while seeming to be weaker based on standard measures of model performance (e.g., AUC, Kappa, PCC), greatly outperformed the spatially representative models based on measures of true model prediction (e.g., correctly describing the actual spatial scales, direction and strength of species-environment relationships). Further work using simulation approaches is warranted to more fully evaluate the ability of species distribution modelling techniques to correctly identify scales, driving covariates, signs and magnitudes of relationships between species presence-absence patterns, and environmental covariates.