Species detection vs. habitat suitability: Are we biasing habitat suitability models with remotely sensed data?

Remotely sensed datasets are increasingly being used to model habitat suitability for a variety of taxa. We review habitat suitability models (HSMs) developed for both plants and animals that include remote sensing predictor variables to determine how these variables could affect model projections. For models focused on plant species habitat, we find several instances of unintentional bias in HSMs of vegetation due to the inclusion of remote sensing variables. Notably, studies that include continuous remote sensing variables could be inadvertently mapping actual species distribution instead of potential habitat due to unique spectral or temporal characteristics of the target species. Additionally, HSMs including categorical classifications are rarely explicit about assumptions of habitat suitability related to land cover, which could lead to unintended exclusion of potential habitat due to current land use. Although we support the broader application of remote sensing in general, we caution developers of HSMs to be aware of introduced model bias. These biases are more likely to arise when remote sensing variables are added to models simply because they improve accuracy, rather than considering how they affect the model results and interpretation. When including land cover classifications as predictors, we recommend that modellers provide more explicit descriptions of how habitat is defined (e.g., is deforested land considered suitable for trees?). Further, we suggest that continuous remote sensing variables should only be included in habitat models if authors can demonstrate that their inclusion characterizes potential habitat rather than actual species distribution. Use of the term ‘habitat suitability model’ rather than ‘species distribution model’ could reduce confusion about modelling goals and improve communication between the remote sensing and ecological modelling communities.