Professional Documents
Culture Documents
Ddi 12970
Ddi 12970
DOI: 10.1111/ddi.12970
BIODIVERSITY RESEARCH
1
College of Fisheries, Ocean University of
China, Qingdao, China Abstract
2
School of Marine Sciences, University of Aim: Biodiversity prediction becomes increasingly important in the face of global
Maine, Orono, ME, USA
diversity loss, whereas substantial challenges still exist in both conceptual and tech‐
3
Qingdao National Laboratory for Marine
Science and Technology, Qingdao, China
nical aspects. There exist many predictive models, and an integrative evaluation can
help understand their performance in handling the multifacets of biodiversity. This
Correspondence
Yiping Ren, College of Fisheries, Ocean
study aims to evaluate the performance of these modelling approaches to predict
University of China, 216, Fisheries Hall, 5 both α‐ and β‐diversity in diverse ecological contexts.
Yushan Road, Qingdao 266003, China.
Email: renyip@ouc.edu.cn
Location: North Yellow Sea, China.
Methods: The biodiversity models follow three strategies, “assemble first, predict
Funding information
National Natural Science Foundation of
later”, “predict first, assemble later” and “assemble and predict together”. Hill diver‐
China, Grant/Award Number: 31772852 and sity profile, Fisher's log‐series parameter and the distance decay of similarity are used
31802301
to measure α‐ and β‐diversity. The evaluation study is conducted based on seasonal
Editor: Eric Treml bottom trawl surveys from October 2016 to August 2017 in North Yellow Sea, China,
allocated to coastal and offshore areas. We evaluate the predictive power of the
models using cross‐validation.
Results: Following the “assemble first, predict later” approach, macroecological
model (MEM) provided the most accurate prediction overall, whereas stacked spe‐
cies distribution model (SSDM) and joint species distribution model (JSDM), following
the second and third modelling approaches, tended to overestimate α‐diversity and
underestimate β‐diversity. The performances of SSDM and JSDM could be improved
by moderately down‐weighting rare species. The relative performances of the three
modelling approaches were consistent among seasons and spatial regions.
Main conclusions: The superior performances of MEM in a range of temporal and
spatial contexts favour the “assemble first, predict later” approach and imply a tight
community assembly in the studied area. The overall predictive powers of varying
models suggest that the spatial pattern of marine biodiversity could be fairly well
predicted with commonly accessible hydrologic data in a mesoscales. The approach
of multi‐model evaluations is applicable to a variety of ecosystems for biodiversity
prediction.
KEYWORDS
GDM, Hill diversity profile, HMSC, Random Forest, SSDM, β‐diversity
This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium,
provided the original work is properly cited.
© 2019 The Authors. Diversity and Distributions Published by John Wiley & Sons Ltd.
1 | I NTRO D U C TI O N D'Amen et al., 2017). In addition, the effects of biotic interaction
may depend on abiotic gradients, as environmental effects often
Biodiversity is a central topic of ecology, and there is a large body dominate communities in stressful areas and species interactions
of literature demonstrating its critical role in ecosystem functions are more influential in benign areas (i.e. the “species interactions–
and services (Isbell et al., 2018; Naeem, Chazdon, Duffy, Prager, & abiotic stress” hypothesis, Louthan, Doak, & Angert, 2015). As
Worm, 2016). From a pragmatic viewpoint, predicting biodiversity such, the driving forces of community assembly depend on the
is critical to identify management priorities and biological conser‐ spatial scale of observations, suitability of environmental condi‐
vation (Certain & Planque, 2015; Coll et al., 2016; Keith, Martin, tions, population density and species' dispersal/migration ability.
McDonald‐Madden, & Walters, 2011; Lindenmayer, Margules, & Therefore, it is unlikely that any one approach to biodiversity mod‐
Botkin, 2000), which is drawing increasing attention as a result of elling and prediction will outperform other approaches in all situa‐
prevalent ecological threats from climate changes, marine pollu‐ tions (Ferrier & Guisan, 2006).
tion, overexploitation and habitat degradation worldwide (Agardy, With respect to the challenges identified above, it is beneficial to
2000; Cheung et al., 2009; Defeo et al., 2017; Magurran, Dornelas, incorporate the strengths of multiple modelling strategies to handle
Moyes, Gotelli, & McGill, 2015; Urban et al., 2016). Although the the multifacets of biodiversity. Although modelling methods of the
topic is appealing, the challenge of biodiversity prediction is daunt‐ three strategies have been well established, they are rarely compared
ing. The concept of biodiversity is complicated, involving richness, and evaluated on the same stage. In addition, most studies have
abundance, evenness and variations, as well as components in the been focused on occurrence data and species richness, and biodi‐
genetic, taxonomic, and functional levels (Jarzyna & Jetz, 2016; versity models are rarely examined based on abundance data (Algar,
Ricotta, 2005). Although there are numerous measurements of di‐ Kharouba, Young, & Kerr, 2009; D'Amen et al., 2017; Santini et al.,
versity (Magurran & McGill, 2011), no single metric can depict the 2017). The present study evaluates the performances of various ex‐
full scope of biodiversity (Santini et al., 2017). Moreover, influencing isting modelling approaches to predict the spatial patterns of α‐ and
factors and driving forces for biodiversity are not always clear and β‐diversity based on abundance data. With respect to the multifacets
observable when measurement errors and ecological stochasticity of biodiversity, multiple metrics are adopted to reflect both α‐ and β‐
largely blur the relationships between biodiversity and predictors, diversity (Jarzyna & Jetz, 2016), and their spatial distributions serve
leading to limited predictive powers in biodiversity models (Iknayan, as the targets of biodiversity prediction. We evaluate the predictive
Tingley, Furnas, & Beissinger, 2014; Jarzyna & Jetz, 2016). Therefore, power of different modelling approaches using cross‐validation and
increasing research efforts are needed for the methodological de‐ compare their performances in different ecological contexts. The
velopment as well as comparing, evaluating and applying biodiver‐ study is implemented in a marine ecosystem in Yellow Sea, China
sity models. based on seasonal surveys, whereas the evaluation approaches are
Many models have been developed to address the challenges applicable to a variety of ecosystems. The objective of this study is
of biodiversity prediction in space. As reviewed by Ferrier and to improve our understanding of community assemblies and provide
Guisan (2006) and D'Amen, Rahbek, Zimmermann, and Guisan guidance for the improvement of biodiversity predictions.
(2017), existing spatial modelling methods can be classified into
three strategies according to their conceptual and statistical
2 | M E TH O DS
bases. The first strategy is to “assemble first, predict later”, in
which species data are first assembled into community‐level attri‐
2.1 | Study area and data
butes, such as species richness, then modelled with environmental
predictors. The second strategy is to “predict first, assemble later”, The study was based on a marine survey in North Yellow Sea (NYS),
in which each species is modelled individually then species‐level China (Figure S1). The surveys were conducted seasonally from
predictions are stacked to community‐level attributes. The third October 2016 to August 2017. A total of 118 sampling sites were
strategy is to “assemble and predict together”, in which all species investigated in each of the four cruises using otter trawls. The sur‐
are modelled simultaneously in an integrated model, and biodi‐ vey area was stratified into three sub‐regions, south coastal area
versity is derived from species‐level predictions accordingly. The A, north coastal area B and offshore area, in which 40, 43, and 35
performance of the three modelling strategies depends on specific stations were sampled, respectively (Figure 1). The trawl had a net
ecological contexts, as ecological drivers of community structure width 15 m and cod‐end mesh 20 mm, hauling for 1 hr at a speed
change substantially in different spatial scales (Rahbek & Graves, 3 knots in each sampling site. Environmental variables including tem‐
2001). For example, environmental filters often take effects in a perature, salinity and depth were measured using CTD (XR‐420) at
large spatial scale by limiting the range of species distributions, each sampling station after hauling.
whereas biotic interactions are more influential in a fine scale via We included 105 nekton species in our analyses after excluding
predation and competition (Araújo & Rozenfeld, 2014; Wisz et al., occasional catches (rare fish or benthos caught with a few individ‐
2013). However, there are also contrary evidences that microcli‐ uals), which might cause the instability of model fitting. It should
mates may affect local communities and biotic interactions influ‐ be noted that removing those rare species could lead to underes‐
ence species distribution in large ranges (Araújo & Luoto, 2007; timation of species richness; however, the influence was limited for
ZHANG et al. |
1699
39
colour denotes the coastal and offshore
areas, and cross denotes the sampling Bohai Sea
sites. The area is divided according to
water depth and management boundary Area B
38
Latitude (N°)
Shandong Peninsula
37
Offshore area
Area A
36
35
Yellow Sea
abundance‐based diversity metrics when evaluating models' relative Species abundance distribution (SAD) is an alternative approach
performances. The abundance of each species that was standard‐ to reflect α‐diversity with the preservation of more information
ized to the same sampling efforts were used as response variables compared to ecological indices (McGill et al., 2007). The SADs de‐
directly or aggregated to diversity metrics before modelling. The scribes the relative abundance (commonness and rarity) of different
water depth, temperature and salinity measured at the sea bottom species, often exhibiting a hyperbolic curve (i.e. few abundant spe‐
at each sampling site were used as the predictive variables for spa‐ cies and many rare species, Figure S5). A range of statistical models
tial biodiversity modelling (environmental data were summarized in have been used to depict the shape of SADs, including log‐series,
Table S1 and Figure S2). Gamma, lognormal, Weibull, Broken stick and zero‐sum multinomial
distribution (Magurran & McGill, 2011). We adopted Fisher's log‐se‐
2.2 | Biodiversity metrics ries model (Fisher, Corbet, & Williams, 1943) to fit abundance data
and used the log‐series parameter as a diversity metric (referred to
Instead of the contemporary focuses on occurrence and species
as f below).
richness, this study used abundance (number of individuals) as the
β‐diversity is often defined as the variation of species composi‐
currency to measure biodiversity (Algar et al., 2009; D'Amen et al.,
tion in a community over space or time (Whittaker, 1960). Among a
2017; Santini et al., 2017). A range of diversity metrics was adopted
wide range of optional metrics (Tuomisto, 2010), this study measures
in this study to reflect both α‐ and β‐diversity. Hill's number was
β‐diversity as the dissimilarity of species composition among sampling
used to depict α‐diversity synthetically (Hill, 1973),
units in different geographical distances, that is, distance decay of simi‐
( )1∕(1−r) larity (Barwell, Isaac, & Kunin, 2015). Two popular metrics, Bray–Curtis
Hr =
∑
pri index (BC) and Cao's similarity index (CY) were used to indicate the dis‐
i similarities in species abundance (Figure S6). The former is a traditional
where pi is the proportion of species i in the sample, and r denotes measure of similarity in ecology (Bray & Curtis, 1957),
the order of diversity. The value of r ranges from zero to infinity, and
∑ � �
an increasing r reduces the weight of rare species (Figure S4). By k �xik − xjk ��
BCij = ∑ �� �
changing r, this equation yields most well‐known diversity indices, k xik + xjk
for example, H0 is simply species richness (when r = 0), H1 equals The latter index, developed by Cao, Williams, and Bark (1997),
the exponential of Shannon diversity index, and H2 is the inverse shows satisfactory performances in measuring β‐diversity (Cao &
Simpson index. Exhibiting different r with the corresponding Hr in a Epifanio, 2010; Barwell et al., 2015).
plot leads to Hill diversity profile, which provides more information
S [ ( ) ]
than any individual diversity index. In this study, the order r (referred 1∑ xik + xjk xik log xjk + xjk log xik
CYij = log −
to as level in this study) ranges from 0 to 4 in the diversity profile S k=1 2 xik + xjk
In both equations, xik is the abundance of species k in sample i, methods in species distribution modelling (Hastie & Tibshirani,
and S is the total number of species. The Bray–Curtis and Cao's index 1990; Guisan, Edwards, & Hastie, 2002), using regression tech‐
were used to illustrate the distance decay of similarity. The Hill di‐ niques, semi‐parametric smoothing functions and various types of
versity profile, Fisher's log‐series model and similarity metrics were error distributions to handle complex species‐environment rela‐
calculated using the “vegan” R package. tionships. In the present study, a two‐step GAM was adopted to
account for the many zeros in multispecies surveys which caused
failures in many modelling techniques, by modelling absence–
2.3 | Modelling approaches
presence data and positive catches separately (O'Neill & Faddy,
We followed the paradigm of Ferrier and Guisan (2006) to adopt 2003; Zhang & Chen, 2015). A alternative Tweedie‐GAM was used
three conceptually different modelling approaches. In the first “as‐ instead when the two steps failed for rare (unfeasible in the sec‐
semble first, predict later” approach, biodiversity metrics were ond step) or prevalent species (unfeasible in the first step) (Candy,
modelled directly with environmental variables. This approach 2004). Spatial coordinates (latitude and longitude) were also in‐
was also referred to as MEM (Algar et al., 2009; Guisan & Rahbek, cluded in modelling to facilitate spatial predictions according to
2011; Certain, Dormann, & Planque, 2014). We adopted two flex‐ preliminary analyses of model predictability (Algar et al., 2009).
ible statistical techniques to tackle possible nonlinear relationships The algorithm was implemented using the “mgcv ” package in R.
between biodiversity and environmental predictors. In the third modelling approach, we adopt a newly developed
The first MEM considered is Random Forest (RF) model for “joint species distribution model” (JSDM) to predict the distri‐
predicting α‐diversity. The RF is a machine‐learning technique bution of all species and their associations altogether (Warton
based on the ensemble of a large number of bootstrapped clas‐ et al., 2015). The JSDMs commonly relate multivariate response
sification and regression trees (Breiman, 2001), in which each variables to abiotic predictors using regression methods (but see
tree is structured by splitting nodes with optimal variables from neural network in Harris, 2015) while accounting for the residual
a random subset of predictors. The prediction of RF is derived as biotic interactions using random effects (Pollock et al., 2014; Hui,
an average over all trees in the forest. The number of variables 2016; Clark, Nemergut, Seyednasrollah, Turner, & Zhang, 2017;
selected at each node (mtry) and the number of trees in a forest Ovaskainen et al., 2017; Thorson & Barnett, 2017; Schliep et al.,
(ntree) needs to be specified in the algorithm (Degenhardt, Seifert, 2018). The random effects are specified by either unstructured
& Szymczak, 2017). Random Forest model is evident to be insensi‐ variance–covariance matrices or unobserved “latent variables”
tive to underlying error distributions, collinearity and interactions (Skrondal & Rabe‐Hesketh, 2004). This study followed the latent‐
among predictors (Di Franco et al., 2016; Duffy, Lefcheck, Stuart‐ variable approach and adopted a JSDM framework “Hierarchical
Smith, Navarrete, & Edgar, 2016). We used spatial coordinates Modelling of Species Communities” (HMSC) (Ovaskainen et al.,
in addition to environmental variables as predictors to facilitate 2017). Hierarchical Modelling of Species Communities uses a
spatial predictions. The second MEM considered is Generalized Bayesian hierarchical method to incorporate environmental vari‐
Dissimilarity Model (GDM) for predicting β‐diversity. The GDM ables, species traits, phylogeny, spatial and temporal structures in
uses the distance in environmental space to explain biological vari‐ community modelling (Ovaskainen & Soininen, 2011). The model
ations (Ferrier, Manion, Elith, & Richardson, 2007). That is, the re‐ uses sparse infinite factor priors for the loadings of latent variables
sponse variable in GDM is a matrix of compositional dissimilarities to account for uncertainty in the number of latent factors and
(e.g. Bray–Curtis similarity index), and the explanatory variables sparsity structure (Bhattacharya & Dunson, 2011). We assumed
are environmental predictors transformed to distance matrices spatial autocorrelation in the structure of latent variables to facili‐
among every pairs of sampling sites (Ferrier et al., 2007; Lasram, tate spatial prediction (Ovaskainen et al., 2017). The algorithm was
Hattab, Halouani, Romdhane, & Le Loc'h, 2015). This algorithm implemented using the “HMSC” package (version 2.1‐1 on GitHub,
is technically an extension of permutation matrix regression that https://github.com/guiblanchet /HMSC).
uses splines to accommodate nonlinearity relationships. The two In summary, the abundance of multiple species were used as the
algorithms were implemented using the “random Forest ” (version response variables in SSDM and JSDM, and the derived diversity
4.6 on CRAN) and “GDM” package (version 1.3.7 on GitHub, https metrics from abundance data were used as the response variables in
://github.com/fitzLab-AL/GDM), respectively. MEM. The same predictive variables were used for different model‐
In the second approach, α‐ and β‐diversity are derived from ling approaches. In addition to hydrologic variables, the latitude and
species‐level predictions, for which the abundance of individual longitude of each sampling site were used to be predictive variables
species is modelled with environmental predictors using spe‐ in MEM and SSDM and to formulate spatial autocorrelation in JSDM.
cies distribution models, such as GLM, GAM and Maxent (Elith
& Leathwick, 2009). This approach is referred to as stacked spe‐
2.4 | Performance evaluation
cies distribution models (SSDM). Generalized additive model
(GAM) was used to build SSDM in this study regarding its flexi‐ We used threefold cross‐validations to examine the predictability
bility in tackling irregular data and nonlinear ecological relation‐ of the three modelling approaches. Two thirds of the sampling sites
ships. Generalized additive model is one of the most widely used were randomly drawn from the survey data as the training set for
ZHANG et al. |
1701
model fitting, and the remaining 1/3 of the data were used as the A, B and offshore, Figure 1) as different ecological contexts for
testing set to evaluate model predictions. Two methods were used model evaluation, and the model fitting and evaluation analyses
to evaluate the accuracy of the predictions on α‐ and β‐diversity. were developed for a combination of seasons and regions sepa‐
A linear regression analysis was conducted between observations rately, resulting in a total of 16,000 sets of evaluation results (4
and model outputs, and the regression slope and the determination seasons × 4 regions × 1000 cross‐validation iterations). The over‐
coefficient were used to describe model accuracy. Normalized root all design of model fitting and evaluation was illustrated in a flow
mean square error (NRMSE), calculated as below, was also used to chart (Figure S3).
measure the divergence of prediction from observation,
√
√ N
√∑ (
√ )2 ( ) 3 | R E S U LT S
NRMSE = x̂ i − xi ∕N∕ xmax − xmin
i
3.1 | α‐diversity
where x̂ i is the predicted value in ith sampling site, and xi was the
observed value in the same sites. xmax and xmin denote the maximum The three modelling approaches (MEM, SSDM, JSDM) showed dif‐
and minimum values in the observation data. Normalized root mean ferent accuracies in the predictions of Fisher's log‐series parameter
square error ranges from 0 to 1 with a lower value indicating better f. According to the linear regression between prediction and ob‐
predictions. servation over the study area (examples shown in Figure S7), MEM
The processes of data splitting, model fitting and accuracy ex‐ showed the highest predictive accuracy, with the regression slope
amination were repeated 1000 times in cross‐validation. The three close to 1.0 and the coefficient of determination (R 2) over 0.9, fol‐
modelling approaches were implemented and examined inde‐ lowed by SSDM (Figure S10, using data in autumn as an example).
pendently, from which we could compare and identify the optimal Fisher's f tended to be underestimated by MEM and largely over‐
strategy for biodiversity prediction. Regarding the context‐de‐ estimated by SSDM and JSDM. The NRMSE also showed accurate
pendent performances of different models (Ferrier & Guisan, predictions by MEM, followed by SSDM. The same analyses were
2006), the cross‐validation was conducted for varying survey sea‐ conducted across the four seasons and the results were generally
sons and regions, respectively. The intensive evaluations aimed consistent, that is, MEM showed superior performances over SSDM
to tackle the seasonality of NYS ecosystem driven by monsoon and JSDM with substantially high precision. For the three model‐
climate. Besides, fishing activities in this region, regulated by an ling approaches, NRMSE ranged from 0.2 to 0.6 (Figure 2, based
annual moratorium policy in summer and adverse climate in win‐ on data of the whole area) and R 2 ranged from 0.8 to 0.95 (Figures
ter, might also contribute to the substantial changes in community S11, S12), indicating an overall satisfactory predictability, although
structure among seasons (Jiang, Cheng, & Li, 2009). In addition, there were remarkable variations of prediction among models and
the biotic community showed significant differences between seasons. Generally, Fisher's f was better predicted in spring and fall
coastal and offshore waters (unpublished data from our survey), compared to that in summer and winter, especially for SSDM.
which should be accounted for in model evaluation. Therefore, we Hill diversity profile provided additional information on the vari‐
treated the four seasons and four regions (whole area, coastal area ation of biodiversity predictions. Macroecological model showed
1.00
NRMSE
0.75
0.50
0.25
Spring Summer Fall Winter Spring Summer Fall Winter Spring Summer Fall Winter
Seasons
F I G U R E 2 The predictive accuracy of Fisher's f by three modelling approaches in the four seasons. The MEM, SSDM and JSDM denote
macroecological model, stacked species distribution model, and joint species distribution model, respectively. The NRMSE denotes the
normalized root mean square error among observation and prediction values
1702 | ZHANG et al.
constantly better performances among the three modelling ap‐ was a constant trend in Bray–Curtis indices that increased with the
2
proaches with respect to regression Slope, R and NRMSE (Figures geographical distances among sampling sites, ranging from 0.60 to
S8, S13 and S14), but underestimated Hill's number in all levels. 0.95, although the similarity varied substantially in similar distances
Stacked species distribution model and JSDM tended to overesti‐ (Figure S9). The prediction of MEM closely matched the observed
mate Hill's number in low levels whereas the bias decreased with patterns of distance decay. Stacked species distribution model and
the increasing levels, minimized at H1 (Figure 3). All modelling ap‐ JSDM predicted less dissimilarity in small distances but proper es‐
proaches tended to underestimate Hill's number when the level r timation in large distances (Figure 4). The relative performances of
was large. Different from the results of Fisher's f, JSDM tended to the three modelling approaches were consistent among the four
outperform SSDM in some situations (e.g. summer in Figure 3). seasons. The similarity decay was nonlinear across the geographical
range of this study, and the rate of decay generally changed at a geo‐
graphical distance of 1.5–2.0 units of the scaled spatial coordinates
3.2 | β‐diversity
(divided by standard deviations). Cao's similarity index showed a sim‐
Both the observations and predictions of β‐diversity showed remark‐ ilar pattern to BC in distance decay but the divergences were more
able distance decay of similarity in community compositions. There remarkable among the three modelling approaches (Figure S15).
Spring Summer
1.5
Regression coefficient between observation and prediction
1.0
0.5
Models
MEM
Fall Winter SSDM
JSDM
1.5
1.0
0.5
0 0.1 0.2 0.3 0.4 0.5 0.8 1 1.5 2 3 4 0 0.1 0.2 0.3 0.4 0.5 0.8 1 1.5 2 3 4
Levels
F I G U R E 3 The predictive accuracy of Hill diversity profile by three modelling approaches in four seasons. The levels in X‐axis indicates
the order of Hill's number. MEM, SSDM and JSDM denote macroecological model, stacked species distribution model and joint species
distribution model, respectively. Slope is the regression coefficient between the observation and prediction of Hill's number
ZHANG et al. |
1703
Bray−Curtis dissimilarity
MEM, SSDM and JSDM denote prediction 0.4
from macroecological model, stacked Models
species distribution model and joint 0.2
MEM
species distribution model, respectively. Fall Winter SSDM
TRUE denotes the observed β‐diversity in JSDM
surveys 1.0 TRUE
0.8
0.6
0.4
0.2
0 1 2 3 4 5 0 1 2 3 4 5
Geographic distance
1.0
Models
NRMSE
MEM
SSDM
JSDM
0.5
0 0.1 0.2 0.3 0.4 0.5 0.8 1 1.5 2 3 4 0 0.1 0.2 0.3 0.4 0.5 0.8 1 1.5 2 3 4 0 0.1 0.2 0.3 0.4 0.5 0.8 1 1.5 2 3 4
Levels
Models
0.8
TRUE
MEM
0.6 SSDM
JSDM
0.4
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
Geographic distance
F I G U R E 5 A comparison of the three modelling approaches in predicting α‐ and β‐diversity in different spatial regions. The subareas
include costal waters (A and B) and offshore. NRMSE denotes the normalized root mean square error between observation and prediction
of Hill diversity profile, and the levels in X‐axis is the order of Hill's number. MEM, SSDM and JSDM denote prediction from macroecological
model, stacked species distribution model and joint species distribution model, respectively
2015), such as the case of biodiversity in climate changes. In fact, uncertainty in deriving biodiversity metrics from species‐level
biodiversity is an emergent result of community assembly, driven predictions. Besides, this modelling approach might also benefit
by diverse processes including biogeographic history, evolutionary from the RF and GDM algorithms, which were evidenced to be
adaptation, environmental filtering, biotic interaction and stochas‐ powerful for prediction among many statistical techniques (Cutler
tic dispersal (Thuiller et al., 2013; Ovaskainen, Abrego, Halme, & et al., 2007; Boulesteix, Janitza, Kruppa, & König, 2012). On the
Dunson, 2016a; D'Amen et al., 2017). Therefore, the transferability other hand, MEM relied on elaborate community‐level inventories
of statistical models is worthy of future investigation. We suggest along environmental gradients, which limited it to predict beyond
that correlative models may be preferable for spatial prediction known communities (Ferrier & Guisan, 2006). As such, SSDM and
of biodiversity in the scope of marine reserves (Parnell, Dayton, JSDM were more flexible for biodiversity extrapolation, when
Lennert‐cody, Rasmussen, & Leichter, 2013), and mechanistic mod‐ ecological niches of individual species were reliable in predic‐
els are necessary to tackle ecosystem shifts under environmental tions (Schmitt, Pouteau, Justeau, Boissieu, & Birnbaum, 2017). It
changes (Singer et al., 2016; Urban et al., 2016; Cabral, Valente, & is worthy to note that their predictions tended to be biased with‐
Hartig, 2017). out considering macroecological constraints (D'Amen et al., 2018),
Comparing performances of multiple modelling approaches for example, overestimating α‐diversity (Guisan & Rahbek, 2011;
contributed to our understanding of their strengths and weak‐ Schmitt et al., 2017). Our results suggested that the performances
nesses, which could serve as a benchmark for model improvement of SSDM and JSDM were likely undermined by rare species, as
(Ferrier & Guisan, 2006; Certain et al., 2014). For example, the both models showed improved performances with increasing lev‐
superior performance of MEM might be attributed to the aggre‐ els in Hill diversity profile which reduced weights on rare species.
gation of species‐level information, that is, modelling collective Besides, rare species also prevented the use of powerful algo‐
properties at the community‐level avoided the challenges of han‐ rithms such as RF and ANN in SSDM, for which only the GAM was
dling individual species especially rare species, in addition to the flexible to avoid fallacious predictions. An extensive survey that
ZHANG et al. |
1705
moderately covered rare species might benefit the predictions of Edwards, Cutler, Zimmermann, Geiser, & Moisen, 2006; Tessarolo,
SSDM and JSDM. Rangel, Araújo, & Hortal, 2014).
Hierarchical Modelling of Species Communities, the algorithm In future studies, the modelling approaches evaluated in this
used in the third strategy, is versatile in handling environmental study can be improved by tuning available model algorithms, while
factors and species interactions in species distribution modelling some metrics may be considered with priority in conservation with
(Ovaskainen et al., 2017). In particular, the inclusion of spatial struc‐ respect to their predictability (e.g. H1), as policymakers often want a
ture of latent variables might greatly improve models' predictive per‐ single number. Meanwhile, additional studies are needed to improve
formances (Ovaskainen, Roy, Fox, & Anderson, 2016b; Ovaskainen the prediction of biodiversity regarding diverse research objectives,
et al., 2017). However, the spatial‐structured HMSC in this study ecological contexts and quality of data, covering the facets of taxo‐
showed no better performances than MEM and SSDM in most sit‐ nomic, functional and phylogenetic components of α‐ and β‐diver‐
uations. The results suggested that environmental gradients domi‐ sity (Mokany, Harwood, Overton, Barker, & Ferrier, 2011; Riera et
nated the spatial pattern of biodiversity in the study area while biotic al., 2018). We suggest that the comparison of multiple predictive
interactions played a minor role (also evidenced in variation partition models and multiple diversity metrics can be useful to provide ad‐
of HMSC). In addition, from a technical viewpoint, HMSC is essen‐ ditional information of biodiversity. However, facing the challenges
tially a multispecies generalized linear mixed model. The algorithm of changes in biodiversity over broad environmental gradients
makes it convenient to model all species simultaneously but con‐ and climate changes, structural integrations of different modelling
strains the capacity to deal with nonlinearity, massive zeros and in‐ modules are desired to combine both mechanistic and correlative
teractions (Zhang, Chen, Xu, Xue, & Ren, 2018), which can be better approaches (Cabral et al., 2017; D'Amen et al., 2017), for example
handled in GAM and RF. Nevertheless, the model' less performance SESAM (Guisan & Rahbek, 2011), Dynamic FOAM (Mokany et al.,
can still help understand the patterns of biodiversity, as the model 2011) and DBEM (Fernandes et al., 2013). Meanwhile, we emphasize
provides evidences of limited biotic interactions which could not be that fully exploring the existing models may be as important as, if no
assured a prior. In addition, the flexible Bayesian hierarchical struc‐ more than, developing new ones. In any case, biodiversity is more
ture (Ovaskainen & Soininen, 2011) may facilitate the improvement than just species number or abundance but should be embedded in
of HMSC in diverse ecological contexts (c.f. Harris, 2015; Thorson & variations, interactions and dynamics, which calls for the develop‐
Barnett, 2017). ment of comprehensive research frameworks.
The choice of diversity metrics and available data are critical for
the predictive power of biodiversity models (Certain et al., 2014;
AC K N OW L E D G E M E N T S
Riera et al., 2018). For example, Certain et al. (2014) suggested that
high predictive performances could be obtained by modelling with Funding for this study was provided by the National Natural Science
species biomass and using diversity measures that give high weight to Foundation of China (31802301, 31772852).
rare species. The conclusion was consistent with the results of MEM
in our study but opposite to that of SSDM and JSDM, which were im‐
proved with increasing r in the Hill's diversity profile (down‐weight‐ DATA AVA I L A B I L I T Y S TAT E M E N T
ing rare species) to certain extents. Our evaluations are conducted in R code and further data supporting the results are provided on
different marine areas (coastal and offshore), and the consistent re‐ Dryad: https://doi.org/10.5061/dryad.2fv7mt8
sults among regions suggest that the superior performance of MEM
may be applicable to diverse ecological contexts. Nevertheless, we
evaluated performances of the modelling approaches based on col‐ ORCID
lective biodiversity properties in this study, and the predictive ability Chongliang Zhang https://orcid.org/0000-0002-4753-9973
could be quite different in terms of individual species distributions.
In addition, the type of available data needs to be considered for
model development regarding different spatial scales. For example, REFERENCES
the “assemble first, predict later” approach relies on local invento‐
Agardy, T. (2000). Effects of fisheries on marine ecosystems: A conser‐
ries of biotic communities that are often obtained at a fine spatial vationist's perspective. ICES Journal of Marine Science, 57, 761–765.
resolution, whereas the “predict first, assemble later” approach https://doi.org/10.1006/jmsc.2000.0721
may be based on species‐occurrence records which are collected Algar, A. C., Kharouba, H. M., Young, E. R., & Kerr, J. T. (2009).
Predicting the future of species diversity: Macroecological
individually at coarse spatial resolutions. Moreover, the relative
theory, climate change, and direct tests of alternative
performances of different models may rely on spatial scales. Given forecasting methods. Ecography, 32, 22–33. https://doi.
the impact of training data on prediction success, future evaluation org/10.1111/j.1600-0587.2009.05832.x
studies of biodiversity models should account for the spatial charac‐ Araújo, M. B., & Luoto, M. (2007). The importance of biotic interac‐
teristics of survey data, including spatial resolution and extent, sur‐ tions for modelling species distributions under climate change.
Global Ecology and Biogeography, 16, 743–753. https://doi.
vey design, spatial autocorrelation, species prevalence, and sample
org/10.1111/j.1466-8238.2007.00359.x
size as well as their interactive effects (Stockwell & Peterson, 2002;
|
1706 ZHANG et al.
Araújo, M. B., & Rozenfeld, A. (2014). The geographic scaling of biotic Ecology and Biogeography, 26, 823–834. https://doi.org/10.1111/
interactions. Ecography, 37, 406–415. geb.12588
Barwell, L. J., Isaac, N. J. B., & Kunin, W. E. (2015). Measuring β‐diver‐ Degenhardt, F., Seifert, S., & Szymczak, S. (2017). Evaluation of variable
sity with species abundance data. Journal of Animal Ecology, 84, selection methods for random forests and omics data sets. Briefings
1112–1122. in Bioinformatics, 20(2), 492–503. https://doi.org/10.1093/bib/
Bhattacharya, A., & Dunson, D. B. (2011). Sparse Bayesian infinite factor bbx124
models. Biometrika, 98, 291–306. https://doi.org/10.1093/biomet/ Di Franco, A., Thiriet, P., Di Carlo, G., Dimitriadis, C., Francour, P.,
asr013 Gutiérrez, N. L., … Guidetti, P. (2016). Five key attributes can increase
Boulesteix, A.‐L., Janitza, S., Kruppa, J., & König, I. R. (2012). Overview marine protected areas performance for small‐scale fisheries man‐
of random forest methodology and practical guidance with emphasis agement. Scientific Reports, 6, 1–9. https://doi.org/10.1038/srep3
on computational biology and bioinformatics. Wiley Interdisciplinary 8135
Reviews: Data Mining and Knowledge Discovery, 2, 493–507. https:// Duffy, J. E., Lefcheck, J. S., Stuart‐Smith, R. D., Navarrete, S. A., & Edgar,
doi.org/10.1002/widm.1072 G. J. (2016). Biodiversity enhances reef fish biomass and resistance to
Bray, J. R., & Curtis, J. T. (1957). An ordination of the upland forest com‐ climate change. Proceedings of the National Academy of Sciences of the
munities of southern Wisconsin. Ecological Monographs, 27, 325–349. United States of America, 113, 6230–6235. https://doi.org/10.1073/
https://doi.org/10.2307/1942268 pnas.1524465113
Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32. Edwards, T. C., Cutler, D. R., Zimmermann, N. E., Geiser, L., & Moisen, G. G.
Cabral, J. S., Valente, L., & Hartig, F. (2017). Mechanistic simulation mod‐ (2006). Effects of sample survey design on the accuracy of classifica‐
els in macroecology and biogeography: State‐of‐art and prospects. tion tree models in species distribution models. Ecological Modelling,
Ecography, 40, 267–280. https://doi.org/10.1111/ecog.02480 199, 132–141. https://doi.org/10.1016/j.ecolmodel.2006.05.016
Candy, S. G. (2004). Modelling catch and effort data using generalised Elith, J., & Leathwick, J. R. (2009). Species distribution models: Ecological
linear models, the Tweedie distribution, random vessel effects and explanation and prediction across space and time. Annual Review
random stratum‐by‐year effects. CCAMLR Science, 11, 59–80. of Ecology, Evolution, and Systematics, 40, 677–697. https://doi.
Cao, Y., & Epifanio, J. (2010). Quantifying the responses of macroinver‐ org/10.1146/annurev.ecolsys.110308.120159
tebrate assemblages to simulated stress: Are more accurate similar‐ Fernandes, J. A., Cheung, W. W. L. L., Jennings, S., Butenschön, M., de
ity indices less useful? Methods in Ecology and Evolution, 1, 380–388. Mora, L., Frölicher, T. L., … Grant, A. (2013). Modelling the effects
https://doi.org/10.1111/j.2041-210X.2010.00040.x of climate change on the distribution and production of marine
Cao, Y., Williams, W. P., & Bark, A. W. (1997). Similarity measure bias fishes: Accounting for trophic interactions in a dynamic bioclimate
in river benthic Aufwuchs community analysis. Water Environment envelope model. Global Change Biology, 19, 2596–2607. https://doi.
Research, 69, 95–106. https://doi.org/10.2175/106143 097X125227 org/10.1111/gcb.12231
Certain, G., Dormann, C. F., & Planque, B. (2014). Choices of abundance Ferrier, S., & Guisan, A. (2006). Spatial modelling of biodiversity at the
currency, community definition and diversity metric control the community level. Journal of Applied Ecology, 43, 393–404. https://doi.
predictive power of macroecological models of biodiversity. Global org/10.1111/j.1365-2664.2006.01149.x
Ecology and Biogeography, 23, 468–478. https://doi.org/10.1111/ Ferrier, S., Manion, G., Elith, J., & Richardson, K. (2007). Using generalized
geb.12119 dissimilarity modelling to analyse and predict patterns of beta diver‐
Certain, G., & Planque, B. (2015). Biodiversity baseline for large marine sity in regional biodiversity assessment. Diversity and Distributions,
ecosystems: An example from the Barents Sea. ICES Journal of Marine 13, 252–264. https://doi.org/10.1111/j.1472-4642.2007.00341.x
Science: Journal Du Conseil, 72, 1756–1768. Fisher, R. A., Corbet, A. S., & Williams, C. B. (1943). The relation between
Cheung, W. W. L., Lam, V. W. Y., Sarmiento, J. L., Kearney, K., Watson, the number of species and the number of individuals in a random
R., & Pauly, D. (2009). Projecting global marine biodiversity impacts sample of an animal population. Journal of Animal Ecology, 12, 42–58.
under climate change scenarios. Fish and Fisheries, 10, 235–251. https://doi.org/10.2307/1411
https://doi.org/10.1111/j.1467-2979.2008.00315.x Guisan, A., Edwards, T. C., & Hastie, T. (2002). Generalized linear and
Clark, J. S., Nemergut, D., Seyednasrollah, B., Turner, P. J., & Zhang, S. generalized additive models in studies of species distributions:
(2017). Generalized joint attribute modeling for biodiversity analysis: Setting the scene. Ecological Modelling, 157, 89–100. https://doi.
Median‐zero, multivariate, multifarious data. Ecological Monographs, org/10.1016/S0304-3800(02)00204-1
87, 34–56. https://doi.org/10.1002/ecm.1241 Guisan, A., & Rahbek, C. (2011). SESAM – A new framework integrating
Coll, M., Shannon, L. J., Kleisner, K. M., Juan‐Jordá, M. J., Bundy, A., macroecological and species distribution models for predicting spa‐
Akoglu, A. G., … Shin, Y.‐J. (2016). Ecological indicators to capture tio‐temporal patterns of species assemblages. Journal of Biogeography,
the effects of fishing on biodiversity and conservation status of 38, 1433–1444. https://doi.org/10.1111/j.1365-2699.2011.02550.x
marine ecosystems. Ecological Indicators, 60, 947–962. https://doi. Harris, D. J. (2015). Generating realistic assemblages with a joint spe‐
org/10.1016/j.ecolind.2015.08.048 cies distribution model. Methods in Ecology and Evolution, 6, 465–473.
Cutler, D. R., Edwards, T. C., Beard, K. H., Cutler, A., Hess, K. T., Gibson, https://doi.org/10.1111/2041-210X.12332
J., & Lawler, J. J. (2007). Random forests for classification in ecology. Hastie, T. J., & Tibshirani, R. J. (1990). Generalized additive models. London,
Ecology, 88, 2783–2792. https://doi.org/10.1890/07-0539.1 UK: Chapman & Hall.
D'Amen, M., Mateo, R. G., Pottier, J., Thuiller, W., Maiorano, L., Pellissier, Hill, M. O. (1973). Diversity and evenness: A unifying notation and its con‐
L., … Guisan, A. (2018). Improving spatial predictions of taxonomic, sequences. Ecology, 54, 427–432. https://doi.org/10.2307/1934352
functional and phylogenetic diversity. Journal of Ecology, 106, 76–86. Hui, F. K. C. (2016). Boral‐Bayesian ordination and regression analysis of
https://doi.org/10.1111/1365-2745.12801 multivariate abundance data in R. Methods in Ecology and Evolution,
D'Amen, M., Rahbek, C., Zimmermann, N. E., & Guisan, A. (2017). Spatial 7, 744–750.
predictions at the community level: From current approaches to Iknayan, K. J., Tingley, M. W., Furnas, B. J., & Beissinger, S. R. (2014).
future frameworks. Biological Reviews, 92, 169–187. https://doi. Detecting diversity: Emerging methods to estimate species diversity.
org/10.1111/brv.12222 Trends in Ecology & Evolution, 29, 97–106. https://doi.org/10.1016/j.
Defeo, O., Barboza, C. A. M., Barboza, F. R., Aeberhard, W. H., Cabrini, tree.2013.10.012
T. M. B., Cardoso, R. S., … Worm, B. (2017). Aggregate patterns Isbell, F., Cowles, J., Dee, L. E., Loreau, M., Reich, P. B., Gonzalez, A., …
of macrofaunal diversity: An interocean comparison. Global Schmid, B. (2018). Quantifying effects of biodiversity on ecosystem
ZHANG et al. |
1707
functioning across times and places. Ecology Letters, 21, 763–778. of community data? A conceptual framework and its implementation
https://doi.org/10.1111/ele.12928 as models and software. Ecology Letters, 20, 561–576. https://doi.
Jarzyna, M. A., & Jetz, W. (2016). Detecting the multiple facets of bio‐ org/10.1111/ele.12757
diversity. Trends in Ecology and Evolution, 31, 527–538. https://doi. Parnell, P. E., Dayton, P. K., Lennert‐cody, C. E., Rasmussen, L. L., & Leichter,
org/10.1016/j.tree.2016.04.002 J. J. (2013). Marine reserve design: Optimal size, habitats, species af‐
Jiang, Y., Cheng, J., & Li, S. (2009). Temporal changes in the fish commu‐ finities, diversity, and ocean microclimate. Ecological Applications, 16,
nity resulting from a summer fishing moratorium in the northern East 945–962. https://doi.org/10.1890/1051-0761(2006)016[0945:MR‐
China Sea. Marine Ecology Progress Series, 387, 265–273. https://doi. DOSH]2.0.CO;2
org/10.3354/meps08 078 Petchey, O. L., Pontarp, M., Massie, T. M., Efi, S. K., Ozgul, A., Weilenmann,
Keith, D. A., Martin, T. G., McDonald‐Madden, E., & Walters, C. (2011). M., … Pearse, I. S. (2015). The ecological forecast horizon, and ex‐
Uncertainty and adaptive management for biodiversity conservation. amples of its uses and determinants. Ecology Letters, 18, 597–611.
Biological Conservation, 144, 1175–1178. https://doi.org/10.1016/j. https://doi.org/10.1111/ele.12443
biocon.2010.11.022 Pollock, L. J., Tingley, R., Morris, W. K., Golding, N., O'Hara, R. B., Parris,
Lasram, F. B. R., Hattab, T., Halouani, G., Romdhane, M. S., & Le Loc'h, K. M., … McCarthy, M. A. (2014). Understanding co‐occurrence by
F. (2015). Modeling of beta diversity in Tunisian waters: Predictions modelling species simultaneously with a Joint Species Distribution
using generalized dissimilarity modeling and bioregionalisation using Model (JSDM). Methods in Ecology and Evolution, 5, 397–406. https://
fuzzy clustering. PLoS ONE, 10, e0131728. https://doi.org/10.1371/ doi.org/10.1111/2041-210X.12180
journal.pone.0131728 Rahbek, C., & Graves, G. R. (2001). Multiscale assessment of patterns
Lindenmayer, D. B., Margules, C. R., & Botkin, D. B. (2000). of avian species richness. Proceedings of the National Academy of
Indicators of biodiversity for ecologically sustainable forest Sciences of the United States of America, 98, 4534–4539. https://doi.
management. Conservation Biology, 14, 941–950. https://doi. org/10.1073/pnas.071034898
org/10.1046/j.1523-1739.2000.98533.x Ricotta, C. (2005). Through the jungle of biological diver‐
Louthan, A. M., Doak, D. F., & Angert, A. L. (2015). Where and when do sity. Acta Biotheoretica, 53, 29–38. https://doi.org/10.1007/
species interactions set range limits? Trends in Ecology and Evolution, s10441-005-7001-6
30, 780–792. https://doi.org/10.1016/j.tree.2015.09.011 Riera, R., Tuset, V. M., Betancur‐R, R., Lombarte, A., Marcos, C., & Pérez‐
Magurran, A. E., Dornelas, M., Moyes, F., Gotelli, N. J., & McGill, B. (2015). Ruzafa, A. (2018). Modelling alpha‐diversities of coastal lagoon fish
Rapid biotic homogenization of marine fish assemblages. Nature assemblages from the Mediterranean Sea. Progress in Oceanography,
Communications, 6, 1–5. https://doi.org/10.1038/ncomms9405 165, 100–109. https://doi.org/10.1016/j.pocean.2018.05.003
Magurran, A. E., & McGill, B. J. (2011). Biological diversity: Frontiers in Santini, L., Belmaker, J., Costello, M. J., Pereira, H. M., Rossberg, A. G.,
measurement and assessment. Oxford, UK: Oxford University Press. Schipper, A. M., … Rondinini, C. (2017). Assessing the suitability of di‐
McGill, B. J., Etienne, R. S., Gray, J. S., Alonso, D., Anderson, M. J., versity metrics to detect biodiversity change. Biological Conservation,
Benecha, H. K., … White, E. P. (2007). Species abundance distribu‐ 213, 341–350. https://doi.org/10.1016/j.biocon.2016.08.024
tions: Moving beyond single prediction theories to integration within Schliep, E. M., Lany, N. K., Zarnetske, P. L., Schaeffer, R. N., Orians, C.
an ecological framework. Ecology Letters, 10, 995–1015. https://doi. M., Orwig, D. A., & Preisser, E. L. (2018). Joint species distribution
org/10.1111/j.1461-0248.2007.01094.x modelling for spatio‐temporal occurrence and ordinal abundance
Mokany, K., Harwood, T. D., Overton, J. M., Barker, G. M., & Ferrier, S. data. Global Ecology and Biogeography, 27, 142–155. https://doi.
(2011). Combining α‐ and β‐diversity models to fill gaps in our knowl‐ org/10.1111/geb.12666
edge of biodiversity. Ecology Letters, 14, 1043–1051. Schmitt, S., Pouteau, R., Justeau, D., de Boissieu, F., & Birnbaum, P.
Mori, A. S., Isbell, F., & Seidl, R. (2018). β‐Diversity, community assem‐ (2017). ssdm: An r package to predict distribution of species rich‐
bly, and ecosystem functioning. Trends in Ecology & Evolution, 1–16. ness and composition based on stacked species distribution models.
https://doi.org/10.1016/j.tree.2018.04.012 Methods in Ecology and Evolution, 8, 1795–1803.
Naeem, S., Chazdon, R., Duffy, J. E., Prager, C., & Worm, B. (2016). Singer, A., Johst, K., Banitz, T., Fowler, M. S., Groeneveld, J., Gutiérrez, A.
Biodiversity and human well‐being: An essential link for sustainable G., … Travis, J. M. J. (2016). Community dynamics under environmen‐
development. Proceedings of the Royal Society B: Biological Sciences, tal change: How can next generation mechanistic models improve
283, 20162091. projections of species distributions? Ecological Modelling, 326, 63–74.
Olden, J. D. (2006). Biotic homogenization: A new research agenda for https://doi.org/10.1016/j.ecolmodel.2015.11.007
conservation biogeography. Journal of Biogeography, 33, 2027–2039. Skrondal, A., & Rabe‐Hesketh, S. (2004). Generalized latent variable mod‐
https://doi.org/10.1111/j.1365-2699.2006.01572.x eling: Multilevel, longitudinal, and structural equation models. New
O'Neill, M. F., & Faddy, M. J. (2003). Use of binary and truncated neg‐ York, NY: Chapman and Hall/CRC.
ative binomial modelling in the analysis of recreational catch Stockwell, D. R. B., & Peterson, A. T. (2002). Effects of sample size on ac‐
data. Fisheries Research, 60, 471–477. https://doi.org/10.1016/ curacy of species distribution models. Ecological Modelling, 148, 1–13.
S0165-7836(02)00101-7 https://doi.org/10.1016/S0304-3800(01)00388-X
Ovaskainen, O., Abrego, N., Halme, P., & Dunson, D. (2016a). Using latent Tessarolo, G., Rangel, T. F., Araújo, M. B., & Hortal, J. (2014). Uncertainty
variable models to identify large networks of species‐to‐species as‐ associated with survey design in Species Distribution Models.
sociations at different spatial scales. Methods in Ecology and Evolution, Diversity and Distributions, 20, 1258–1269. https://doi.org/10.1111/
7, 549–555. https://doi.org/10.1111/2041-210X.12501 ddi.12236
Ovaskainen, O., Roy, D. B., Fox, R., & Anderson, B. J. (2016b). Uncovering Thorson, J. T., & Barnett, L. A. K. (2017). Comparing estimates of abun‐
hidden spatial structure in species communities with spatially explicit dance trends and distribution shifts using single‐ and multispecies
joint species distribution models. Methods in Ecology and Evolution, 7, models of fishes and biogenic habitat. ICES Journal of Marine Science:
428–436. https://doi.org/10.1111/2041-210X.12502 Journal Du Conseil, 74, 1311–1321. https://doi.org/10.1093/icesjms/
Ovaskainen, O., & Soininen, J. (2011). Making more out of sparse data: fsw193
Hierarchical modeling of species communities. Ecology, 92, 289–295. Thuiller, W., Münkemüller, T., Lavergne, S., Mouillot, D., Mouquet, N.,
https://doi.org/10.1890/10-1251.1 Schiffers, K., & Gravel, D. (2013). A road map for integrating eco‐
Ovaskainen, O., Tikhonov, G., Norberg, A., Guillaume Blanchet, F., evolutionary processes into biodiversity models. Ecology Letters, 16,
Duan, L., Dunson, D., … Abrego, N. (2017). How to make more out 94–105. https://doi.org/10.1111/ele.12104
|
1708 ZHANG et al.