Dealing with overprediction in species distribution models: how adding distance constraints can improve model accuracy

Abstract

Species distribution models can be affected by overprediction when dispersal movement is not incorporated into the modelling process. We compared the efficiency of seven methods that take into account spatial constraints to reduce overprediction when using four algorithms for species distribution models. By using a virtual ecologist approach, we were able to measure the accuracy of each model in predicting actual species distributions. We built 40 virtual species distributions within the Neotropical realm. Then, we randomly sampled 50 occurrences that were used in seven spatially restricted species distribution models (hereafter called M-SDMs) and a non-spatially restricted ecological niche model (ENM). We used four algorithms; Maximum Entropy, Generalized Linear Models, Random Forest, and Support Vector Machine. M-SDM methods were divided into a priori methods, in which spatial restrictions were inserted with environmental variables in the modelling process, and a posteriori methods, in which reachable and suitable areas were overlapped. M-SDM efficiency was obtained by calculating the difference in commission and omission errors between M-SDMs and ENMs. We used linear mixed-effects models to test if differences in commission and omission errors varied among the M-SDMs and algorithms. Our results indicate that overall M-SDMs reduce overprediction with no increase in underprediction compared to ENMs with few exceptions, such as a priori methods combined with the Support Vector Machine algorithm. There is a high variation in modelling performance among species, but there were only a few cases in which overprediction or underprediction increased. We only compared methods that do not require species dispersal data, guaranteeing that they can be applied to less-studied species. We advocate that species distribution modellers should not ignore spatial constraints, especially because they can be included in models at low costs but high benefits in terms of overprediction reduction.