Spatially nested species distribution models (N-SDM): An effective tool to overcome niche truncation for more robust inference and projections

1 SETTING THE SCENE

Species distribution models (SDMs), also known as ecological niche models (ENMs) or habitat suitability models (HSMs), among other terms (Elith & Leathwick, 2009; Franklin, 2010; Guisan et al., 2017; Peterson et al., 2011), have become major tools in ecology, conservation and biogeography (Araujo et al., 2019; Ferrier et al., 2016; Guisan et al., 2013). SDMs have a wide array of applications, a key one being to derive spatial projections of species distributions across space and time, that is, to predict where suitable locations will be under future (Guisan & Thuiller, 2005; Patiño et al., 2023; Peterson et al., 2018; Petitpierre et al., 2016) or have been under past (Bruni et al., 2024; Maiorano et al., 2013; Nogues-Bravo, 2009; Pearman et al., 2008) climates, and/or in different geographic areas (Gallien et al., 2012; Guisan et al., 2013; Petitpierre et al., 2017).

SDMs are based on the concept of Hutchinson (1957) realized environmental niche (also known as ecological niche; Austin & Smith, 1989; Guisan et al., 2017; Peterson et al., 2011). The realized niche was formalized as a hypervolume of species requirements in a multidimensional environmental space (Austin et al., 1990; Guisan & Zimmermann, 2000) and is a subset of the fundamental environmental niche constrained by biotic interactions and dispersal limitations (historic factors in Hutchinson's terms; or accessibility in Soberon, 2007). An SDM fitted with empirical observations (e.g. presence–absence, presence-only or abundance data) therefore captures the realized environmental niche—hereafter ecological niche—of the modelled species (Araujo & Guisan, 2006; Guisan & Zimmermann, 2000; Pearman et al., 2008; Pearson & Dawson, 2003; Soberon, 2007).

The validity of SDMs is based on several important assumptions (Anderson, 2013; Franklin, 2010; Guisan et al., 2017; Peterson et al., 2011; Zurell et al., 2020). Among these, a critical one is that the model must capture the entire ecological niche to be appropriately projected to other spatiotemporal contexts (Guisan et al., 2017). When this is not the case, there is a risk that the modelled response curves of the species along environmental gradients are truncated, resulting in biased and often incorrect predictions of species distributions (Chevalier et al., 2021; Figure 1). This phenomenon is known as ‘niche truncation’ and can occur when the geographic extent used to fit the model does not include all possible conditions that compose the ecological niche of a species (Figure 1; Anderson, 2013; Bazzichetto et al., 2023; Pearson et al., 2004; Thuiller, Brotons, et al., 2004). However, not all cases of geographic restriction result in niche truncation because a subset of the species range could cover the entire range of environmental conditions experienced by a species. This problem has been largely overlooked in the extensive SDM literature of the last three decades, despite early studies showing that truncated SDM can lead to biased projections of species distributions under climate change (Box 1).

FIGURE 1

Open in figure viewerPowerPoint

Illustrating the problem of niche truncation by using a training area (Switzerland, CH) smaller than the species geographic range in Europe (EU). Black dots (a) and black curves (c–h) correspond to the full range of species (IUCN range map). Red dots (b) and red curves (c–h) correspond to the restricted training area (CH), and associated SDM predictions. The maps (i and j) show predictions along a gradient from suitable in blue to unsuitable environments in yellow (through green, intermediate suitability), obtained from a model fitted at the EU scale, and thus encompassing the whole species range (i) versus a model fitted on the restricted range only (j). The distribution of unsuitable yellow areas in the two maps (compared to the actual distribution in a) shows that the global model (i) captures much better the distribution of this species than the regional model (j). Graphs (c–h) show, for six environmental predictors in the models, how the response curves fitted with the full range (black dots in a) and restricted range (red dots in b) diverge. Inspired by Chevalier et al. (2021).

BOX 1. A short history on the problem of niche truncation by geographic restriction

The challenge of niche truncation due to geographic restriction has been acknowledged in the scientific community for more than two decades, yet it has not been thoroughly examined until recently. The importance of capturing the entire environmental range of species in SDMs to generate meaningful spatial predictions was first emphasized two decades ago (Pearson et al., 2004; Pearson & Dawson, 2003; Thuiller, Brotons, et al., 2004). Using generalized additive models with varying sizes of nested training ranges for three tree species, Thuiller, Brotons, et al. (2004) showed that models based on restricted ranges produced significantly altered response curves and spatial overpredictions compared to full range models (e.g. erroneously predicting a mediterranean oak species in Scandinavia), leading to an underestimation of species extinction risk under climate change (see also Barbet-Massin et al., 2010 for birds in Europe). The same year, Pearson and co-authors proposed a methodological solution to this problem by fitting SDMs at two nested spatial scales, continental (Europe) to fit the entire climatic niche, and national (Britain) to include more specific land cover habitat preferences (Pearson et al., 2004). Although the original intent of the latter study was to illustrate the effects of distinct environmental variables at different spatial scales (i.e. climate in Europe, land use in the United Kingdom), it also provided a solution to the problem of niche truncation.

In a more theoretical article delving into the factors influencing climate change projections based on SDMs, Anderson (2013) underscored the critical significance of the study region that encompasses the entire spectrum of abiotic conditions a species can inhabit. This is crucial for the abiotic variables considered, as failing to do so means that SDMs can only estimate a portion of the environmental niche of the species. Anderson's work highlighted the issues that arise when using such a restricted model, particularly in terms of the inaccuracies in the environmental response curves of the species when making future projections (Fig. 3 in Anderson, 2013). Owens et al. (2013) addressed the same problem, while expanding on the associated problem of model transferability and extrapolation to novel conditions.

Hannemann et al. (2016) further explored the potential causes of prediction errors in SDMs fitted with truncated training datasets (here Germany only) for seven tree species in Europe, revealing several issues with truncated SDMs, including spurious response curves, inconsistent variable selection and large overprediction of the species across Europe. More recently, Scherrer et al. (2021) compared three scales—global, national and local—to fit SDMs for three Mexican tree species, to assess the effects of a geographic restriction of training data on the estimated environmental niche and on its projection in space and time, also assessing how it can affect the potential vulnerability of species to climate change. The results showed that the effects were species-specific and strongly dependent on the extent to which the geographic truncation affected estimations of environmental niches, with cases of strong niche underestimation leading to more vulnerability to climate change (Scherrer et al., 2021). The same year, Chevalier et al. (2021) illustrated the niche truncation problem using virtual simulations and one real plant species, proposing multiscale data fusion methods as solutions.

Since then, other studies have further discussed this niche truncation issue and its impact on species distribution models (e.g. Adde et al., 2023; Carrillo-García et al., 2023; Chevalier et al., 2022; Goicolea et al., 2024), a topic increasingly attracting attention.

A promising strategy to address the challenges of niche truncation, that is, capturing the full ecological niche while still predicting species distributions at a fine resolution, is to fit two or more spatially nested SDMs (hereafter N-SDMs, as used in Adde et al., 2023; see dedicated section below) covering different extents and resolutions, and combine their predictions. The N-SDM approach usually includes at least (i) a ‘whole range’ SDM that spans the entire species range (often at a global or continental scale) to capture the full species' ecological niche, albeit at a coarse resolution and with a limited number of predictors (typically climatic variables that reflect broad-scale climatic requirements), and (ii) a ‘subrange’ SDM that focuses on a restricted area of interest (e.g. for conservation planning), often containing only a portion of the species range (e.g. at national or regional scale). This subrange model is typically fitted with high-resolution predictors to account for species' detailed, fine-scale requirements, such as habitat or landscape characteristics (Adde et al., 2023; Chevalier et al., 2021, 2022; Goicolea et al., 2024; Mateo et al., 2024; Mateo, Aroca-Fernández, et al., 2019; Riva et al., 2024).

The aim of this review is to present N-SDMs as an effective solution to provide fine-resolution projections at national, regional or local scales while overcoming the problem of niche truncation (Box 1). We review existing developments in this field, synthesize how N-SDMs can combine data at different scales, discuss remaining limitations and challenges, and propose some future perspectives. While doing this, we show that the ongoing development of N-SDMs has the potential to be transformative for biodiversity science. To further clarify our scope, we note that (i) previous studies have used the term ‘niche truncation’ referring to other issues that must not be confused with the one associated here with geographic restriction, and (ii) the implications for climate change projections based on SDMs affected by geographic restriction are closely intertwined with studies on SDM transferability in time and space (Box 2).

BOX 2. Clarifying the concepts of niche truncation and model transferability

To clarify the scope of this review, there are two important comments to make about niche truncation and model transferability.

First, other studies have used the term ‘niche truncation’ to refer to other issues, which should not be confused with the one associated with geographic restriction causing niche truncation that is addressed in this review. For example, the term ‘truncation’ has been used to refer to how much the realized environmental niche is a truncated subset of the fundamental niche (Bush et al., 2018; Chevalier et al., 2024; Vetaas, 2002; Webber et al., 2011), how the realized niche is truncated by the accessible (i.e. dispersal limitations; Peterson et al., 2018) or available environment (Chevalier et al., 2024), or how much humans have caused truncation of current species ranges and niches compared to past ones (Pang et al., 2022; Sales et al., 2022). Although these topics are still related to environmental niche dynamics (Pearman et al., 2008) and can also affect SDM predictions, they are distinct issues that are not addressed in this review.

Second, the implications for climate change projections based on SDMs affected by geographic restriction are closely intertwined with studies on SDM transferability in time and space (Barbosa et al., 2009; Charney et al., 2021; Qiao et al., 2019; Randin et al., 2006; Regos et al., 2019; Yates et al., 2018), where predictions are altered in non-analog conditions (Fitzpatrick & Hargrove, 2009; Petitpierre et al., 2017). Predicting the future distribution of a species from a model affected by niche truncation only makes the transferability issue more complex (Anderson, 2013; Peterson et al., 2018), because (i) such a model does not account for the full conditions actually experienced by the species throughout its entire range and (ii) a larger proportion of non-analog conditions is likely to occur in the future leading to prediction errors also in locations that are included within the training area but where no prediction errors are expected under current conditions (Elith et al., 2010; Velazco et al., 2024). However, this issue embraces a scope far larger than the issue related to niche truncation by geographic restriction (i.e. also touching to other issues around model fitting and evaluation; Peterson et al., 2018; Petitpierre et al., 2017; Yates et al., 2018) and is not addressed in this review.