Importance and effectiveness of correction methods for spatial sampling bias in species with sex-specific habitat preference.

Anouk Glad,Jörn Pagel,Jean‐Matthieu Monnet,Björn Reineking

doi:10.1002/ece3.5765

Anouk Glad, Jörn Pagel + Show 2 more

Open Access

https://doi.org/10.1002/ece3.5765

Copy DOI

Abstract

AimPresence records from surveys with spatially heterogeneous sampling intensity are a key challenge for species distribution models (SDMs). When sex groups differ in their habitat association, the correction of the spatial bias becomes important for preventing model predictions that are biased toward one sex. The objectives of this study were to investigate the effectiveness of existing correction methods for spatial sampling bias for SDMs when male and female have different habitat preferences.LocationJura massif, France.MethodsWe used a spatially sex‐segregated virtual species to understand the effect of three sampling designs (spatially biased, uniform random, and systematic), and two correction methods (targeted background points, and distance to trajectories) on estimated habitat preferences, sex ratios, and prediction accuracy. We then evaluated these effects for two empirical Capercaillie (Tetrao urogallus) presence‐only datasets from a systematic and a spatially biased sampling design.ResultsSampling design strongly affected parameter estimation accuracy for the virtual species: noncorrected spatially biased sampling resulted in biased estimates of habitat association and sex ratios. Both established methods of bias correction were successful in the case of virtual species, with the targeted correction methods showing stronger correction, as it more closely followed the simulated decay of detectability with distance from sampling locations. On the Capercaillie dataset, only the targeted background points method resulted in the same sex ratio estimate for the spatially biased sampling design as for the spatially unbiased sampling.Main conclusionsWe suggest that information on subgroups with distinct habitat associations should be included in SDMs analyses when possible. We conclude that current methods for correcting spatially biased sampling can improve estimates of both habitat association and subgroup ratios (e.g., sex and age), but that their efficiency depends on their ability to well represent the spatial observation bias.

Highlights

IntroductionThe impact of human activities on the environment at different scales, from local (urbanization and agriculture) to global (climate change), necessitates to better understand the relationships be‐ tween species and their environment, and to better predict their distribution
The impact of human activities on the environment at different scales, from local to global, necessitates to better understand the relationships be‐ tween species and their environment, and to better predict their distribution
We evaluated the effect of two correction methods in the case of uneven spatial sampling on estimates of habitat prefer‐ ences, sex ratio estimation, and prediction accuracy

Summary

Introduction

The impact of human activities on the environment at different scales, from local (urbanization and agriculture) to global (climate change), necessitates to better understand the relationships be‐ tween species and their environment, and to better predict their distribution. Occurrence data for distribution modeling come from many sources and often do not originate from controlled sampling survey designs such as systematic transects or random plots, but from spa‐ tially preferential sampling or opportunistic observations (Geldmann et al, 2016). Data collected without a sampling framework are sub‐ ject to sampling bias (Guisan, Thuiller, & Zimmermann, 2017), which can be of different sources: uneven record intensity in space or time and uneven sampling effort and variations in detection efficiency among observers (Geldmann et al, 2016). We focus on the effects of uneven record intensity in space, since spatially biased sampling is a major cause of poor model predictions accuracy (Araújo & Guisan, 2006; Guisan & Zimmermann, 2000; Renner et al, 2015). The effect of sampling bias on prediction may often be underestimated, as many published results are not corrected for sampling bias before analysis (Yackulic et al, 2013)

Objectives

Methods

Findings

Discussion

Conclusion