Data quantity is more important than its spatial bias for predictive species distribution modelling

Abstract

Biological records are often the data of choice for training predictive species distribution models (SDMs), but spatial sampling bias is pervasive in biological records data at multiple spatial scales and is thought to impair the performance of SDMs. We simulated presences and absences of virtual species as well as the process of recording these species to evaluate the effect on species distribution model prediction performance of 1) spatial bias in training data, 2) sample size (the average number of observations per species), and 3) the choice of species distribution modelling method. Our approach is novel in quantifying and applying real-world spatial sampling biases to simulated data. Spatial bias in training data decreased species distribution model prediction performance, but only when the bias was relatively strong. Sample size and the choice of modelling method were more important than spatial bias in determining the prediction performance of species distribution models.

Embracing Ensemble Species Distribution Models to Inform At-Risk Species Status Assessments

“…Given that most of our predictor variables had 30- m resolution and some had 1-km resolution (i.e., Worldclim variables), we decided to use 500 m as the compromise distance to apply this spatial filter. We also used this distance to retain more of the few occurrences available to develop SDMs (Gaul et al 2020). The final number of presence records per species was B. atropurpurea; n ¼ 67, M. caroliniana: n ¼ 115, P. eryngii: n ¼ 54, and S. ocmulgee: n ¼ 22 (Figure 2).…”

“…Lastly, practitioners might question the reliability of models built with a limited number of locations. Indeed, despite considerable advancements in tools and techniques, the efficacy of SDMs lies in the quality of data used to build the models, making presence data a major limiting factor (Phillips et al 2009, Gaul et al 2020). Conversely, SDMs can produce spatial predictions that may help field biologists target new survey areas (Fois et al 2018).…”

Modeling present and future distribution of plankton populations in a coastal upwelling zone: the copepod Calanus chilensis as a study case

“…Therefore, given the presence of bias in the data, the most suitable analysis strategy is the use of approximations that allow this bias to be reduced, reducing its influence on the parameter estimates 70 . Another difficulty when modeling the distribution of endemic species such as C. chilensis is obtaining enough occurrences [72][73][74] , which is a recurrent situation for species that are rare, endemic or with biased sampling 75 . To take into account the low number of occurrences for the 200-400 m strata, we used ESMs, an approach described to date as the most suitable for getting robust predictions even when modeling rare species or with a reduced number of occurrences 76,77 .…”

Modelling the distribution of rare invertebrates by correcting class imbalance and spatial bias

“…In contrast, spatially stratified under‐sampling reduces the spatial bias in non‐detection data, while leaving unchanged the spatial bias in detection data (Robinson et al, 2018). Given the minimal impact of spatial sampling bias on SDMs when the biases are similar in detection and non‐detection data (Gaul et al, 2020; Johnston et al, 2020; Thibaud et al, 2014), and the potential negative impact of having spatial bias in detection data that differs from the bias in non‐detection data (Phillips et al, 2009), it seems possible that manipulating the spatial bias in non‐detection data but not in detection data during spatially stratified under‐sampling might make SDMs worse, not better. Robinson et al (2018) compared SDMs trained with raw and with spatially stratified under‐sampled data, but did not test models trained with data that were under‐sampled in a spatially unstratified way.…”