| |

Received: 21 June 2018    Revised: 13 May 2019    Accepted: 4 July 2019

DOI: 10.1111/ddi.12970

BIODIVERSITY RESEARCH

How to predict biodiversity in space? An evaluation of

modelling approaches in marine ecosystems

Chongliang Zhang1  | Yong Chen2 | Binduo Xu1 | Ying Xue1 | Yiping Ren1,3

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.

Diversity and Distributions. 2019;25:1697–1708. wileyonlinelibrary.com/journal/ddi |  1697

  
 |
1698       ZHANG et al.
1 |  I NTRO D U C TI O N
Biodiversity is a central topic of ecology, and there is a large body
of literature demonstrating its critical role in ecosystem functions
and services (Isbell et al., 2018; Naeem, Chazdon, Duffy, Prager, &
Worm, 2016). From a pragmatic viewpoint, predicting biodiversity
is critical to identify management priorities and biological conser‐
vation (Certain & Planque, 2015; Coll et al., 2016; Keith, Martin,
McDonald‐Madden, & Walters, 2011; Lindenmayer, Margules, &
Botkin, 2000), which is drawing increasing attention as a result of
prevalent ecological threats from climate changes, marine pollu‐
tion, overexploitation and habitat degradation worldwide (Agardy,
2000; Cheung et al., 2009; Defeo et al., 2017; Magurran, Dornelas,
Moyes, Gotelli, & McGill, 2015; Urban et al., 2016). Although the
topic is appealing, the challenge of biodiversity prediction is daunt‐
ing. The concept of biodiversity is complicated, involving richness,
abundance, evenness and variations, as well as components in the
genetic, taxonomic, and functional levels (Jarzyna & Jetz, 2016;
Ricotta, 2005). Although there are numerous measurements of di‐
versity (Magurran & McGill, 2011), no single metric can depict the
full scope of biodiversity (Santini et al., 2017). Moreover, influencing
factors and driving forces for biodiversity are not always clear and
observable when measurement errors and ecological stochasticity
largely blur the relationships between biodiversity and predictors,
leading to limited predictive powers in biodiversity models (Iknayan,
Tingley, Furnas, & Beissinger, 2014; Jarzyna & Jetz, 2016). Therefore,
increasing research efforts are needed for the methodological de‐
velopment as well as comparing, evaluating and applying biodiver‐
sity models.
Many models have been developed to address the challenges
of biodiversity prediction in space. As reviewed by Ferrier and
Guisan (2006) and D'Amen, Rahbek, Zimmermann, and Guisan
(2017), existing spatial modelling methods can be classified into
three strategies according to their conceptual and statistical
bases. The first strategy is to “assemble first, predict later”, in
which species data are first assembled into community‐level attri‐
butes, such as species richness, then modelled with environmental
predictors. The second strategy is to “predict first, assemble later”,
in which each species is modelled individually then species‐level
predictions are stacked to community‐level attributes. The third
strategy is to “assemble and predict together”, in which all species
are modelled simultaneously in an integrated model, and biodi‐
versity is derived from species‐level predictions accordingly. The
performance of the three modelling strategies depends on specific
ecological contexts, as ecological drivers of community structure
change substantially in different spatial scales (Rahbek & Graves,
2001). For example, environmental filters often take effects in a
large spatial scale by limiting the range of species distributions,
whereas biotic interactions are more influential in a fine scale via
predation and competition (Araújo & Rozenfeld, 2014; Wisz et al.,
2013). However, there are also contrary evidences that microcli‐
mates may affect local communities and biotic interactions influ‐
ence species distribution in large ranges (Araújo & Luoto, 2007;
D'Amen et al., 2017). In addition, the effects of biotic interaction
may depend on abiotic gradients, as environmental effects often
dominate communities in stressful areas and species interactions
are more influential in benign areas (i.e. the “species interactions–
abiotic stress” hypothesis, Louthan, Doak, & Angert, 2015). As
such, the driving forces of community assembly depend on the
spatial scale of observations, suitability of environmental condi‐
tions, population density and species' dispersal/migration ability.
Therefore, it is unlikely that any one approach to biodiversity mod‐
elling and prediction will outperform other approaches in all situa‐
tions (Ferrier & Guisan, 2006).
With respect to the challenges identified above, it is beneficial to
incorporate the strengths of multiple modelling strategies to handle
the multifacets of biodiversity. Although modelling methods of the
three strategies have been well established, they are rarely compared
and evaluated on the same stage. In addition, most studies have
been focused on occurrence data and species richness, and biodi‐
versity models are rarely examined based on abundance data (Algar,
Kharouba, Young, & Kerr, 2009; D'Amen et al., 2017; Santini et al.,
2017). The present study evaluates the performances of various ex‐
isting modelling approaches to predict the spatial patterns of α‐ and
β‐diversity based on abundance data. With respect to the multifacets
of biodiversity, multiple metrics are adopted to reflect both α‐ and β‐
diversity (Jarzyna & Jetz, 2016), and their spatial distributions serve
as the targets of biodiversity prediction. We evaluate the predictive
power of different modelling approaches using cross‐validation and
compare their performances in different ecological contexts. The
study is implemented in a marine ecosystem in Yellow Sea, China
based on seasonal surveys, whereas the evaluation approaches are
applicable to a variety of ecosystems. The objective of this study is
to improve our understanding of community assemblies and provide
guidance for the improvement of biodiversity predictions.
2 | M E TH O DS
2.1 | Study area and data
The study was based on a marine survey in North Yellow Sea (NYS),
China (Figure S1). The surveys were conducted seasonally from
October 2016 to August 2017. A total of 118 sampling sites were
investigated in each of the four cruises using otter trawls. The sur‐
vey area was stratified into three sub‐regions, south coastal area
A, north coastal area B and offshore area, in which 40, 43, and 35
stations were sampled, respectively (Figure 1). The trawl had a net
width 15 m and cod‐end mesh 20 mm, hauling for 1 hr at a speed
3 knots in each sampling site. Environmental variables including tem‐
perature, salinity and depth were measured using CTD (XR‐420) at
each sampling station after hauling.
We included 105 nekton species in our analyses after excluding
occasional catches (rare fish or benthos caught with a few individ‐
uals), which might cause the instability of model fitting. It should
be noted that removing those rare species could lead to underes‐
timation of species richness; however, the influence was limited for
ZHANG et al. |
      1699

F I G U R E 1   Survey areas and sampling

sites in North Yellow Sea (NYS). The

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

119 120 121 122 123 124 125

Longitude (E°)

abundance‐based diversity metrics when evaluating models' relative
performances. The abundance of each species that was standard‐
ized to the same sampling efforts were used as response variables
directly or aggregated to diversity metrics before modelling. The
water depth, temperature and salinity measured at the sea bottom
at each sampling site were used as the predictive variables for spa‐
tial biodiversity modelling (environmental data were summarized in
Table S1 and Figure S2).
2.2 | Biodiversity metrics
Instead of the contemporary focuses on occurrence and species
richness, this study used abundance (number of individuals) as the
currency to measure biodiversity (Algar et al., 2009; D'Amen et al.,
2017; Santini et al., 2017). A range of diversity metrics was adopted
in this study to reflect both α‐ and β‐diversity. Hill's number was
used to depict α‐diversity synthetically (Hill, 1973),
( )1∕(1−r)
Species abundance distribution (SAD) is an alternative approach
to reflect α‐diversity with the preservation of more information
compared to ecological indices (McGill et al., 2007). The SADs de‐
scribes the relative abundance (commonness and rarity) of different
species, often exhibiting a hyperbolic curve (i.e. few abundant spe‐
cies and many rare species, Figure S5). A range of statistical models
have been used to depict the shape of SADs, including log‐series,
Gamma, lognormal, Weibull, Broken stick and zero‐sum multinomial
distribution (Magurran & McGill, 2011). We adopted Fisher's log‐se‐
ries model (Fisher, Corbet, & Williams, 1943) to fit abundance data
and used the log‐series parameter as a diversity metric (referred to
as f below).
β‐diversity is often defined as the variation of species composi‐
tion in a community over space or time (Whittaker, 1960). Among a
wide range of optional metrics (Tuomisto, 2010), this study measures
β‐diversity as the dissimilarity of species composition among sampling
units in different geographical distances, that is, distance decay of simi‐
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

(Certain & Planque, 2015).

 |
1700       ZHANG et al.
In both equations, xik is the abundance of species k in sample i,
and S is the total number of species. The Bray–Curtis and Cao's index
were used to illustrate the distance decay of similarity. The Hill di‐
versity profile, Fisher's log‐series model and similarity metrics were
calculated using the “vegan” R package.
2.3 | Modelling approaches
We followed the paradigm of Ferrier and Guisan (2006) to adopt
three conceptually different modelling approaches. In the first “as‐
semble first, predict later” approach, biodiversity metrics were
modelled directly with environmental variables. This approach
was also referred to as MEM (Algar et al., 2009; Guisan & Rahbek,
2011; Certain, Dormann, & Planque, 2014). We adopted two flex‐
ible statistical techniques to tackle possible nonlinear relationships
between biodiversity and environmental predictors.
The first MEM considered is Random Forest (RF) model for
predicting α‐diversity. The RF is a machine‐learning technique
based on the ensemble of a large number of bootstrapped clas‐
sification and regression trees (Breiman, 2001), in which each
tree is structured by splitting nodes with optimal variables from
a random subset of predictors. The prediction of RF is derived as
an average over all trees in the forest. The number of variables
selected at each node (mtry) and the number of trees in a forest
(ntree) needs to be specified in the algorithm (Degenhardt, Seifert,
& Szymczak, 2017). Random Forest model is evident to be insensi‐
tive to underlying error distributions, collinearity and interactions
among predictors (Di Franco et al., 2016; Duffy, Lefcheck, Stuart‐
Smith, Navarrete, & Edgar, 2016). We used spatial coordinates
in addition to environmental variables as predictors to facilitate
spatial predictions. The second MEM considered is Generalized
Dissimilarity Model (GDM) for predicting β‐diversity. The GDM
uses the distance in environmental space to explain biological vari‐
ations (Ferrier, Manion, Elith, & Richardson, 2007). That is, the re‐
sponse variable in GDM is a matrix of compositional dissimilarities
(e.g. Bray–Curtis similarity index), and the explanatory variables
are environmental predictors transformed to distance matrices
among every pairs of sampling sites (Ferrier et al., 2007; Lasram,
Hattab, Halouani, Romdhane, & Le Loc'h, 2015). This algorithm
is technically an extension of permutation matrix regression that
uses splines to accommodate nonlinearity relationships. The two
algorithms were implemented using the “random Forest ” (version
4.6 on CRAN) and “GDM” package (version 1.3.7 on GitHub, https​
://github.com/fitzL​ab-AL/GDM), respectively.
In the second approach, α‐ and β‐diversity are derived from
species‐level predictions, for which the abundance of individual
species is modelled with environmental predictors using spe‐
cies distribution models, such as GLM, GAM and Maxent (Elith
& Leathwick, 2009). This approach is referred to as stacked spe‐
cies distribution models (SSDM). Generalized additive model
(GAM) was used to build SSDM in this study regarding its flexi‐
bility in tackling irregular data and nonlinear ecological relation‐
ships. Generalized additive model is one of the most widely used
methods in species distribution modelling (Hastie & Tibshirani,
1990; Guisan, Edwards, & Hastie, 2002), using regression tech‐
niques, semi‐parametric smoothing functions and various types of
error distributions to handle complex species‐environment rela‐
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–
presence data and positive catches separately (O'Neill & Faddy,
2003; Zhang & Chen, 2015). A alternative Tweedie‐GAM was used
instead when the two steps failed for rare (unfeasible in the sec‐
ond step) or prevalent species (unfeasible in the first step) (Candy,
2004). Spatial coordinates (latitude and longitude) were also in‐
cluded in modelling to facilitate spatial predictions according to
preliminary analyses of model predictability (Algar et al., 2009).
The algorithm was implemented using the “mgcv ” package in R.
In the third modelling approach, we adopt a newly developed
“joint species distribution model” (JSDM) to predict the distri‐
bution of all species and their associations altogether (Warton
et al., 2015). The JSDMs commonly relate multivariate response
variables to abiotic predictors using regression methods (but see
neural network in Harris, 2015) while accounting for the residual
biotic interactions using random effects (Pollock et al., 2014; Hui,
2016; Clark, Nemergut, Seyednasrollah, Turner, & Zhang, 2017;
Ovaskainen et al., 2017; Thorson & Barnett, 2017; Schliep et al.,
2018). The random effects are specified by either unstructured
variance–covariance matrices or unobserved “latent variables”
(Skrondal & Rabe‐Hesketh, 2004). This study followed the latent‐
variable approach and adopted a JSDM framework “Hierarchical
Modelling of Species Communities” (HMSC) (Ovaskainen et al.,
2017). Hierarchical Modelling of Species Communities uses a
Bayesian hierarchical method to incorporate environmental vari‐
ables, species traits, phylogeny, spatial and temporal structures in
community modelling (Ovaskainen & Soininen, 2011). The model
uses sparse infinite factor priors for the loadings of latent variables
to account for uncertainty in the number of latent factors and
sparsity structure (Bhattacharya & Dunson, 2011). We assumed
spatial autocorrelation in the structure of latent variables to facili‐
tate spatial prediction (Ovaskainen et al., 2017). The algorithm was
implemented using the “HMSC” package (version 2.1‐1 on GitHub,
https​://github.com/guibl​anche​t /HMSC).
In summary, the abundance of multiple species were used as the
response variables in SSDM and JSDM, and the derived diversity
metrics from abundance data were used as the response variables in
MEM. The same predictive variables were used for different model‐
ling approaches. In addition to hydrologic variables, the latitude and
longitude of each sampling site were used to be predictive variables
in MEM and SSDM and to formulate spatial autocorrelation in JSDM.
2.4 | Performance evaluation
We used threefold cross‐validations to examine the predictability
of the three modelling approaches. Two thirds of the sampling sites
were randomly drawn from the survey data as the training set for
ZHANG et al.
model fitting, and the remaining 1/3 of the data were used as the
testing set to evaluate model predictions. Two methods were used
to evaluate the accuracy of the predictions on α‐ and β‐diversity.
A linear regression analysis was conducted between observations
and model outputs, and the regression slope and the determination
coefficient were used to describe model accuracy. Normalized root
mean square error (NRMSE), calculated as below, was also used to
measure the divergence of prediction from observation,
√
√ N
√∑ (
√ )2 ( ) 3 | R E S U LT S
NRMSE = x̂ i − xi ∕N∕ xmax − xmin
i
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
and minimum values in the observation data. Normalized root mean
square error ranges from 0 to 1 with a lower value indicating better
predictions.
The processes of data splitting, model fitting and accuracy ex‐
amination were repeated 1000 times in cross‐validation. The three
modelling approaches were implemented and examined inde‐
pendently, from which we could compare and identify the optimal
strategy for biodiversity prediction. Regarding the context‐de‐
pendent performances of different models (Ferrier & Guisan,
2006), the cross‐validation was conducted for varying survey sea‐
sons and regions, respectively. The intensive evaluations aimed
to tackle the seasonality of NYS ecosystem driven by monsoon
climate. Besides, fishing activities in this region, regulated by an
annual moratorium policy in summer and adverse climate in win‐
ter, might also contribute to the substantial changes in community
structure among seasons (Jiang, Cheng, & Li, 2009). In addition,
the biotic community showed significant differences between
coastal and offshore waters (unpublished data from our survey),
which should be accounted for in model evaluation. Therefore, we
treated the four seasons and four regions (whole area, coastal area
MEM
1.25
1.00
NRMSE
0.75
0.50
0.25
Spring Summer Fall Winter Spring Summer
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
|
      1701
A, B and offshore, Figure 1) as different ecological contexts for
model evaluation, and the model fitting and evaluation analyses
were developed for a combination of seasons and regions sepa‐
rately, resulting in a total of 16,000 sets of evaluation results (4
seasons × 4 regions × 1000 cross‐validation iterations). The over‐
all design of model fitting and evaluation was illustrated in a flow
chart (Figure S3).
3.1 | α‐diversity
The three modelling approaches (MEM, SSDM, JSDM) showed dif‐
ferent accuracies in the predictions of Fisher's log‐series parameter
f. According to the linear regression between prediction and ob‐
servation over the study area (examples shown in Figure S7), MEM
showed the highest predictive accuracy, with the regression slope
close to 1.0 and the coefficient of determination (R 2) over 0.9, fol‐
lowed by SSDM (Figure S10, using data in autumn as an example).
Fisher's f tended to be underestimated by MEM and largely over‐
estimated by SSDM and JSDM. The NRMSE also showed accurate
predictions by MEM, followed by SSDM. The same analyses were
conducted across the four seasons and the results were generally
consistent, that is, MEM showed superior performances over SSDM
and JSDM with substantially high precision. For the three model‐
ling approaches, NRMSE ranged from 0.2 to 0.6 (Figure 2, based
on data of the whole area) and R 2 ranged from 0.8 to 0.95 (Figures
S11, S12), indicating an overall satisfactory predictability, although
there were remarkable variations of prediction among models and
seasons. Generally, Fisher's f was better predicted in spring and fall
compared to that in summer and winter, especially for SSDM.
Hill diversity profile provided additional information on the vari‐
ation of biodiversity predictions. Macroecological model showed
SSDM JSDM
Fall Winter Spring Summer Fall Winter
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

F I G U R E 4   The prediction of Spring Summer

β‐diversity by three modelling approaches
in four seasons. Distance decay of 1.0

similarity is measured as Bray–Curtis

index along geographical distance among 0.8
sampling sites. The distance in the X‐axis
is calculated from the scaled spatial
0.6
coordinates, and the lines denote the
loess smooth of all pairs of sampling sites.

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

changes. As a step towards an integrative research framework,

3.3 | Regional evaluation
Model performances in predicting α‐ and β‐diversity were evalu‐
ated in sub‐regions of the study area, including two coastal areas
(A and B) and the offshore area. The superior performance of
MEM was consistent with the results in the whole area, whereas
the three modelling approaches showed larger divergences in their
performances (Figure 5). In the south coastal area A, MEM showed
increasing NRMSE with the increasing levels of Hill's number,
whereas NRMSE firstly decreased then stabilized in SSDM and
JSDM. In the north coastal area B, NRMSE was relatively stable
in MEM and tended to decease in the other models except the
hump of JSDM at the level of 0.4. The pattern of Hill profile in
offshore area was similar to area A, and SSDM and JSDM yielded
close predictions.
The MEM also provided better predictions of β‐diversity in the
subareas, whereas SSDM and JSDM substantially underestimated
β‐diversity in short distances (Figure 5). The community dissimilarity
ranged from 0.5 to 0.8 in coastal areas, substantially less than that of
the offshore area. There was a large change of similarity decay rates
at the geographical distance of 1.6 units (spatial coordinates scaled
by standard deviations) in the offshore area, captured by all three
models. Such a pattern was obscure in the coastal areas.
4 | D I S CU S S I O N
Predicting the changes of biodiversity has received growing at‐
tention given increasing concerns about habitat and climate
this study evaluated a range of modelling approaches to predict
the spatial pattern of α‐ and β‐diversity based on abundance data.
Particularly, abundance‐based diversity and β‐diversity should
be highlighted regarding the requirement of multi‐functionality
and ongoing biotic homogenization of marine ecosystems (Olden,
2006; Magurran et al., 2015) and the prevalent research focuses
on species richness (Mori, Isbell, & Seidl, 2018). In general, our
study showed that the three modelling approaches evaluated
in this study had fair predictive power on biodiversity, implying
that marine biodiversity could be reasonably predicted with com‐
monly accessible hydrologic data (i.e. temperature, salinity and
depth), although the driving forces behind biodiversity are com‐
plex. Additionally, MEM provided most accurate predictions in a
range of temporal and spatial contexts, whereas SSDM and JSDM
tended to overestimate α‐diversity and underestimate β‐diversity
in certain circumstances. Our results favoured the “assemble first,
predict later” modelling approach. The study was conducted in the
specific areas of NYS, whereas the consistent results among sea‐
sons and regions suggested the modelling approaches might be
applicable to diverse ecological contexts for spatial prediction of
biodiversity.
Although the results were encouraging, it should be noted that
the cross‐validation of this study was optimistic, as the predictions
were interpolated into the same area of model fitting where underly‐
ing ecological relationships tended to be consistent. In other words,
the modelling methods were based on correlative relationships
which might be less reliable when model fitting and prediction were
conducted in different areas and/or temporal scales (Petchey et al.,
1704       | ZHANG et al.

Area A Area B Offshore

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

Area A Area B Offshore

1.0
Bray−Curtis dissimilarity

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,
biodiversity is an emergent result of community assembly, driven
by diverse processes including biogeographic history, evolutionary
adaptation, environmental filtering, biotic interaction and stochas‐
tic dispersal (Thuiller et al., 2013; Ovaskainen, Abrego, Halme, &
Dunson, 2016a; D'Amen et al., 2017). Therefore, the transferability
of statistical models is worthy of future investigation. We suggest
that correlative models may be preferable for spatial prediction
of biodiversity in the scope of marine reserves (Parnell, Dayton,
Lennert‐cody, Rasmussen, & Leichter, 2013), and mechanistic mod‐
els are necessary to tackle ecosystem shifts under environmental
changes (Singer et al., 2016; Urban et al., 2016; Cabral, Valente, &
Hartig, 2017).
Comparing performances of multiple modelling approaches
contributed to our understanding of their strengths and weak‐
nesses, which could serve as a benchmark for model improvement
(Ferrier & Guisan, 2006; Certain et al., 2014). For example, the
superior performance of MEM might be attributed to the aggre‐
gation of species‐level information, that is, modelling collective
properties at the community‐level avoided the challenges of han‐
dling individual species especially rare species, in addition to the
uncertainty in deriving biodiversity metrics from species‐level
predictions. Besides, this modelling approach might also benefit
from the RF and GDM algorithms, which were evidenced to be
powerful for prediction among many statistical techniques (Cutler
et al., 2007; Boulesteix, Janitza, Kruppa, & König, 2012). On the
other hand, MEM relied on elaborate community‐level inventories
along environmental gradients, which limited it to predict beyond
known communities (Ferrier & Guisan, 2006). As such, SSDM and
JSDM were more flexible for biodiversity extrapolation, when
ecological niches of individual species were reliable in predic‐
tions (Schmitt, Pouteau, Justeau, Boissieu, & Birnbaum, 2017). It
is worthy to note that their predictions tended to be biased with‐
out considering macroecological constraints (D'Amen et al., 2018),
for example, overestimating α‐diversity (Guisan & Rahbek, 2011;
Schmitt et al., 2017). Our results suggested that the performances
of SSDM and JSDM were likely undermined by rare species, as
both models showed improved performances with increasing lev‐
els in Hill diversity profile which reduced weights on rare species.
Besides, rare species also prevented the use of powerful algo‐
rithms such as RF and ANN in SSDM, for which only the GAM was
flexible to avoid fallacious predictions. An extensive survey that
ZHANG et al.
moderately covered rare species might benefit the predictions of
SSDM and JSDM.
Hierarchical Modelling of Species Communities, the algorithm
used in the third strategy, is versatile in handling environmental
factors and species interactions in species distribution modelling
(Ovaskainen et al., 2017). In particular, the inclusion of spatial struc‐
ture of latent variables might greatly improve models' predictive per‐
formances (Ovaskainen, Roy, Fox, & Anderson, 2016b; Ovaskainen
et al., 2017). However, the spatial‐structured HMSC in this study
showed no better performances than MEM and SSDM in most sit‐
uations. The results suggested that environmental gradients domi‐
nated the spatial pattern of biodiversity in the study area while biotic
interactions played a minor role (also evidenced in variation partition
of HMSC). In addition, from a technical viewpoint, HMSC is essen‐
tially a multispecies generalized linear mixed model. The algorithm
makes it convenient to model all species simultaneously but con‐
strains the capacity to deal with nonlinearity, massive zeros and in‐
teractions (Zhang, Chen, Xu, Xue, & Ren, 2018), which can be better
handled in GAM and RF. Nevertheless, the model' less performance
can still help understand the patterns of biodiversity, as the model
provides evidences of limited biotic interactions which could not be
assured a prior. In addition, the flexible Bayesian hierarchical struc‐
ture (Ovaskainen & Soininen, 2011) may facilitate the improvement
of HMSC in diverse ecological contexts (c.f. Harris, 2015; Thorson &
Barnett, 2017).
The choice of diversity metrics and available data are critical for
the predictive power of biodiversity models (Certain et al., 2014;
Riera et al., 2018). For example, Certain et al. (2014) suggested that
high predictive performances could be obtained by modelling with
species biomass and using diversity measures that give high weight to
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‐
ing rare species) to certain extents. Our evaluations are conducted in
different marine areas (coastal and offshore), and the consistent re‐
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‐
lective biodiversity properties in this study, and the predictive ability
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,
the “assemble first, predict later” approach relies on local invento‐
ries of biotic communities that are often obtained at a fine spatial
resolution, whereas the “predict first, assemble later” approach
may be based on species‐occurrence records which are collected
individually at coarse spatial resolutions. Moreover, the relative
performances of different models may rely on spatial scales. Given
the impact of training data on prediction success, future evaluation
studies of biodiversity models should account for the spatial charac‐
teristics of survey data, including spatial resolution and extent, sur‐
vey design, spatial autocorrelation, species prevalence, and sample
size as well as their interactive effects (Stockwell & Peterson, 2002;
|
      1705
Edwards, Cutler, Zimmermann, Geiser, & Moisen, 2006; Tessarolo,
Rangel, Araújo, & Hortal, 2014).
In future studies, the modelling approaches evaluated in this
study can be improved by tuning available model algorithms, while
some metrics may be considered with priority in conservation with
respect to their predictability (e.g. H1), as policymakers often want a
single number. Meanwhile, additional studies are needed to improve
the prediction of biodiversity regarding diverse research objectives,
ecological contexts and quality of data, covering the facets of taxo‐
nomic, functional and phylogenetic components of α‐ and β‐diver‐
sity (Mokany, Harwood, Overton, Barker, & Ferrier, 2011; Riera et
al., 2018). We suggest that the comparison of multiple predictive
models and multiple diversity metrics can be useful to provide ad‐
ditional information of biodiversity. However, facing the challenges
of changes in biodiversity over broad environmental gradients
and climate changes, structural integrations of different modelling
modules are desired to combine both mechanistic and correlative
approaches (Cabral et al., 2017; D'Amen et al., 2017), for example
SESAM (Guisan & Rahbek, 2011), Dynamic FOAM (Mokany et al.,
2011) and DBEM (Fernandes et al., 2013). Meanwhile, we emphasize
that fully exploring the existing models may be as important as, if no
more than, developing new ones. In any case, biodiversity is more
than just species number or abundance but should be embedded in
variations, interactions and dynamics, which calls for the develop‐
ment of comprehensive research frameworks.
AC K N OW L E D G E M E N T S
Funding for this study was provided by the National Natural Science
Foundation of China (31802301, 31772852).
DATA AVA I L A B I L I T Y S TAT E M E N T
R code and further data supporting the results are provided on
Dryad: https​://doi.org/10.5061/dryad.2fv7mt8
ORCID
Chongliang Zhang  https://orcid.org/0000-0002-4753-9973
REFERENCES
Agardy, T. (2000). Effects of fisheries on marine ecosystems: A conser‐
vationist's perspective. ICES Journal of Marine Science, 57, 761–765.
https​://doi.org/10.1006/jmsc.2000.0721
Algar, A. C., Kharouba, H. M., Young, E. R., & Kerr, J. T. (2009).
Predicting the future of species diversity: Macroecological
theory, climate change, and direct tests of alternative
forecasting methods. Ecography, 32, 22–33. https​://doi.
org/10.1111/j.1600-0587.2009.05832.x
Araújo, M. B., & Luoto, M. (2007). The importance of biotic interac‐
tions for modelling species distributions under climate change.
Global Ecology and Biogeography, 16, 743–753. https​://doi.
org/10.1111/j.1466-8238.2007.00359.x
 |
1706       ZHANG et al.
Araújo, M. B., & Rozenfeld, A. (2014). The geographic scaling of biotic
interactions. Ecography, 37, 406–415.
Barwell, L. J., Isaac, N. J. B., & Kunin, W. E. (2015). Measuring β‐diver‐
sity with species abundance data. Journal of Animal Ecology, 84,
1112–1122.
Bhattacharya, A., & Dunson, D. B. (2011). Sparse Bayesian infinite factor
models. Biometrika, 98, 291–306. https​://doi.org/10.1093/biome​t/
asr013
Boulesteix, A.‐L., Janitza, S., Kruppa, J., & König, I. R. (2012). Overview
of random forest methodology and practical guidance with emphasis
on computational biology and bioinformatics. Wiley Interdisciplinary
Reviews: Data Mining and Knowledge Discovery, 2, 493–507. https​://
doi.org/10.1002/widm.1072
Bray, J. R., & Curtis, J. T. (1957). An ordination of the upland forest com‐
munities of southern Wisconsin. Ecological Monographs, 27, 325–349.
https​://doi.org/10.2307/1942268
Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32.
Cabral, J. S., Valente, L., & Hartig, F. (2017). Mechanistic simulation mod‐
els in macroecology and biogeography: State‐of‐art and prospects.
Ecography, 40, 267–280. https​://doi.org/10.1111/ecog.02480​
Candy, S. G. (2004). Modelling catch and effort data using generalised
linear models, the Tweedie distribution, random vessel effects and
random stratum‐by‐year effects. CCAMLR Science, 11, 59–80.
Cao, Y., & Epifanio, J. (2010). Quantifying the responses of macroinver‐
tebrate assemblages to simulated stress: Are more accurate similar‐
ity indices less useful? Methods in Ecology and Evolution, 1, 380–388.
https​://doi.org/10.1111/j.2041-210X.2010.00040.x
Cao, Y., Williams, W. P., & Bark, A. W. (1997). Similarity measure bias
in river benthic Aufwuchs community analysis. Water Environment
Research, 69, 95–106. https​://doi.org/10.2175/10614​3 097X​125227
Certain, G., Dormann, C. F., & Planque, B. (2014). Choices of abundance
currency, community definition and diversity metric control the
predictive power of macroecological models of biodiversity. Global
Ecology and Biogeography, 23, 468–478. https​://doi.org/10.1111/
geb.12119​
Certain, G., & Planque, B. (2015). Biodiversity baseline for large marine
ecosystems: An example from the Barents Sea. ICES Journal of Marine
Science: Journal Du Conseil, 72, 1756–1768.
Cheung, W. W. L., Lam, V. W. Y., Sarmiento, J. L., Kearney, K., Watson,
R., & Pauly, D. (2009). Projecting global marine biodiversity impacts
under climate change scenarios. Fish and Fisheries, 10, 235–251.
https​://doi.org/10.1111/j.1467-2979.2008.00315.x
Clark, J. S., Nemergut, D., Seyednasrollah, B., Turner, P. J., & Zhang, S.
(2017). Generalized joint attribute modeling for biodiversity analysis:
Median‐zero, multivariate, multifarious data. Ecological Monographs,
87, 34–56. https​://doi.org/10.1002/ecm.1241
Coll, M., Shannon, L. J., Kleisner, K. M., Juan‐Jordá, M. J., Bundy, A.,
Akoglu, A. G., … Shin, Y.‐J. (2016). Ecological indicators to capture
the effects of fishing on biodiversity and conservation status of
marine ecosystems. Ecological Indicators, 60, 947–962. https​://doi.
org/10.1016/j.ecoli​nd.2015.08.048
Cutler, D. R., Edwards, T. C., Beard, K. H., Cutler, A., Hess, K. T., Gibson,
J., & Lawler, J. J. (2007). Random forests for classification in ecology.
Ecology, 88, 2783–2792. https​://doi.org/10.1890/07-0539.1
D'Amen, M., Mateo, R. G., Pottier, J., Thuiller, W., Maiorano, L., Pellissier,
L., … Guisan, A. (2018). Improving spatial predictions of taxonomic,
functional and phylogenetic diversity. Journal of Ecology, 106, 76–86.
https​://doi.org/10.1111/1365-2745.12801​
D'Amen, M., Rahbek, C., Zimmermann, N. E., & Guisan, A. (2017). Spatial
predictions at the community level: From current approaches to
future frameworks. Biological Reviews, 92, 169–187. https​://doi.
org/10.1111/brv.12222​
Defeo, O., Barboza, C. A. M., Barboza, F. R., Aeberhard, W. H., Cabrini,
T. M. B., Cardoso, R. S., … Worm, B. (2017). Aggregate patterns
of macrofaunal diversity: An interocean comparison. Global
Ecology and Biogeography, 26, 823–834. https​://doi.org/10.1111/
geb.12588​
Degenhardt, F., Seifert, S., & Szymczak, S. (2017). Evaluation of variable
selection methods for random forests and omics data sets. Briefings
in Bioinformatics, 20(2), 492–503. https​://doi.org/10.1093/bib/
bbx124
Di Franco, A., Thiriet, P., Di Carlo, G., Dimitriadis, C., Francour, P.,
Gutiérrez, N. L., … Guidetti, P. (2016). Five key attributes can increase
marine protected areas performance for small‐scale fisheries man‐
agement. Scientific Reports, 6, 1–9. https​://doi.org/10.1038/srep3​
8135
Duffy, J. E., Lefcheck, J. S., Stuart‐Smith, R. D., Navarrete, S. A., & Edgar,
G. J. (2016). Biodiversity enhances reef fish biomass and resistance to
climate change. Proceedings of the National Academy of Sciences of the
United States of America, 113, 6230–6235. https​://doi.org/10.1073/
pnas.15244​65113​
Edwards, T. C., Cutler, D. R., Zimmermann, N. E., Geiser, L., & Moisen, G. G.
(2006). Effects of sample survey design on the accuracy of classifica‐
tion tree models in species distribution models. Ecological Modelling,
199, 132–141. https​://doi.org/10.1016/j.ecolm​odel.2006.05.016
Elith, J., & Leathwick, J. R. (2009). Species distribution models: Ecological
explanation and prediction across space and time. Annual Review
of Ecology, Evolution, and Systematics, 40, 677–697. https​://doi.
org/10.1146/annur​ev.ecols​ys.110308.120159
Fernandes, J. A., Cheung, W. W. L. L., Jennings, S., Butenschön, M., de
Mora, L., Frölicher, T. L., … Grant, A. (2013). Modelling the effects
of climate change on the distribution and production of marine
fishes: Accounting for trophic interactions in a dynamic bioclimate
envelope model. Global Change Biology, 19, 2596–2607. https​://doi.
org/10.1111/gcb.12231​
Ferrier, S., & Guisan, A. (2006). Spatial modelling of biodiversity at the
community level. Journal of Applied Ecology, 43, 393–404. https​://doi.
org/10.1111/j.1365-2664.2006.01149.x
Ferrier, S., Manion, G., Elith, J., & Richardson, K. (2007). Using generalized
dissimilarity modelling to analyse and predict patterns of beta diver‐
sity in regional biodiversity assessment. Diversity and Distributions,
13, 252–264. https​://doi.org/10.1111/j.1472-4642.2007.00341.x
Fisher, R. A., Corbet, A. S., & Williams, C. B. (1943). The relation between
the number of species and the number of individuals in a random
sample of an animal population. Journal of Animal Ecology, 12, 42–58.
https​://doi.org/10.2307/1411
Guisan, A., Edwards, T. C., & Hastie, T. (2002). Generalized linear and
generalized additive models in studies of species distributions:
Setting the scene. Ecological Modelling, 157, 89–100. https​://doi.
org/10.1016/S0304-3800(02)00204-1
Guisan, A., & Rahbek, C. (2011). SESAM – A new framework integrating
macroecological and species distribution models for predicting spa‐
tio‐temporal patterns of species assemblages. Journal of Biogeography,
38, 1433–1444. https​://doi.org/10.1111/j.1365-2699.2011.02550.x
Harris, D. J. (2015). Generating realistic assemblages with a joint spe‐
cies distribution model. Methods in Ecology and Evolution, 6, 465–473.
https​://doi.org/10.1111/2041-210X.12332​
Hastie, T. J., & Tibshirani, R. J. (1990). Generalized additive models. London,
UK: Chapman & Hall.
Hill, M. O. (1973). Diversity and evenness: A unifying notation and its con‐
sequences. Ecology, 54, 427–432. https​://doi.org/10.2307/1934352
Hui, F. K. C. (2016). Boral‐Bayesian ordination and regression analysis of
multivariate abundance data in R. Methods in Ecology and Evolution,
7, 744–750.
Iknayan, K. J., Tingley, M. W., Furnas, B. J., & Beissinger, S. R. (2014).
Detecting diversity: Emerging methods to estimate species diversity.
Trends in Ecology & Evolution, 29, 97–106. https​://doi.org/10.1016/j.
tree.2013.10.012
Isbell, F., Cowles, J., Dee, L. E., Loreau, M., Reich, P. B., Gonzalez, A., …
Schmid, B. (2018). Quantifying effects of biodiversity on ecosystem
ZHANG et al.
functioning across times and places. Ecology Letters, 21, 763–778.
https​://doi.org/10.1111/ele.12928​
Jarzyna, M. A., & Jetz, W. (2016). Detecting the multiple facets of bio‐
diversity. Trends in Ecology and Evolution, 31, 527–538. https​://doi.
org/10.1016/j.tree.2016.04.002
Jiang, Y., Cheng, J., & Li, S. (2009). Temporal changes in the fish commu‐
nity resulting from a summer fishing moratorium in the northern East
China Sea. Marine Ecology Progress Series, 387, 265–273. https​://doi.
org/10.3354/meps0​8 078
Keith, D. A., Martin, T. G., McDonald‐Madden, E., & Walters, C. (2011).
Uncertainty and adaptive management for biodiversity conservation.
Biological Conservation, 144, 1175–1178. https​://doi.org/10.1016/j.
biocon.2010.11.022
Lasram, F. B. R., Hattab, T., Halouani, G., Romdhane, M. S., & Le Loc'h,
F. (2015). Modeling of beta diversity in Tunisian waters: Predictions
using generalized dissimilarity modeling and bioregionalisation using
fuzzy clustering. PLoS ONE, 10, e0131728. https​://doi.org/10.1371/
journ​al.pone.0131728
Lindenmayer, D. B., Margules, C. R., & Botkin, D. B. (2000).
Indicators of biodiversity for ecologically sustainable forest
management. Conservation Biology, 14, 941–950. https​://doi.
org/10.1046/j.1523-1739.2000.98533.x
Louthan, A. M., Doak, D. F., & Angert, A. L. (2015). Where and when do
species interactions set range limits? Trends in Ecology and Evolution,
30, 780–792. https​://doi.org/10.1016/j.tree.2015.09.011
Magurran, A. E., Dornelas, M., Moyes, F., Gotelli, N. J., & McGill, B. (2015).
Rapid biotic homogenization of marine fish assemblages. Nature
Communications, 6, 1–5. https​://doi.org/10.1038/ncomm​s9405​
Magurran, A. E., & McGill, B. J. (2011). Biological diversity: Frontiers in
measurement and assessment. Oxford, UK: Oxford University Press.
McGill, B. J., Etienne, R. S., Gray, J. S., Alonso, D., Anderson, M. J.,
Benecha, H. K., … White, E. P. (2007). Species abundance distribu‐
tions: Moving beyond single prediction theories to integration within
an ecological framework. Ecology Letters, 10, 995–1015. https​://doi.
org/10.1111/j.1461-0248.2007.01094.x
Mokany, K., Harwood, T. D., Overton, J. M., Barker, G. M., & Ferrier, S.
(2011). Combining α‐ and β‐diversity models to fill gaps in our knowl‐
edge of biodiversity. Ecology Letters, 14, 1043–1051.
Mori, A. S., Isbell, F., & Seidl, R. (2018). β‐Diversity, community assem‐
bly, and ecosystem functioning. Trends in Ecology & Evolution, 1–16.
https​://doi.org/10.1016/j.tree.2018.04.012
Naeem, S., Chazdon, R., Duffy, J. E., Prager, C., & Worm, B. (2016).
Biodiversity and human well‐being: An essential link for sustainable
development. Proceedings of the Royal Society B: Biological Sciences,
283, 20162091.
Olden, J. D. (2006). Biotic homogenization: A new research agenda for
conservation biogeography. Journal of Biogeography, 33, 2027–2039.
https​://doi.org/10.1111/j.1365-2699.2006.01572.x
O'Neill, M. F., & Faddy, M. J. (2003). Use of binary and truncated neg‐
ative binomial modelling in the analysis of recreational catch
data. Fisheries Research, 60, 471–477. https​://doi.org/10.1016/
S0165-7836(02)00101-7
Ovaskainen, O., Abrego, N., Halme, P., & Dunson, D. (2016a). Using latent
variable models to identify large networks of species‐to‐species as‐
sociations at different spatial scales. Methods in Ecology and Evolution,
7, 549–555. https​://doi.org/10.1111/2041-210X.12501​
Ovaskainen, O., Roy, D. B., Fox, R., & Anderson, B. J. (2016b). Uncovering
hidden spatial structure in species communities with spatially explicit
joint species distribution models. Methods in Ecology and Evolution, 7,
428–436. https​://doi.org/10.1111/2041-210X.12502​
Ovaskainen, O., & Soininen, J. (2011). Making more out of sparse data:
Hierarchical modeling of species communities. Ecology, 92, 289–295.
https​://doi.org/10.1890/10-1251.1
Ovaskainen, O., Tikhonov, G., Norberg, A., Guillaume Blanchet, F.,
Duan, L., Dunson, D., … Abrego, N. (2017). How to make more out
|
      1707
of community data? A conceptual framework and its implementation
as models and software. Ecology Letters, 20, 561–576. https​://doi.
org/10.1111/ele.12757​
Parnell, P. E., Dayton, P. K., Lennert‐cody, C. E., Rasmussen, L. L., & Leichter,
J. J. (2013). Marine reserve design: Optimal size, habitats, species af‐
finities, diversity, and ocean microclimate. Ecological Applications, 16,
945–962. https​://doi.org/10.1890/1051-0761(2006)016[0945:MR‐
DOS​H]2.0.CO;2
Petchey, O. L., Pontarp, M., Massie, T. M., Efi, S. K., Ozgul, A., Weilenmann,
M., … Pearse, I. S. (2015). The ecological forecast horizon, and ex‐
amples of its uses and determinants. Ecology Letters, 18, 597–611.
https​://doi.org/10.1111/ele.12443​
Pollock, L. J., Tingley, R., Morris, W. K., Golding, N., O'Hara, R. B., Parris,
K. M., … McCarthy, M. A. (2014). Understanding co‐occurrence by
modelling species simultaneously with a Joint Species Distribution
Model (JSDM). Methods in Ecology and Evolution, 5, 397–406. https​://
doi.org/10.1111/2041-210X.12180​
Rahbek, C., & Graves, G. R. (2001). Multiscale assessment of patterns
of avian species richness. Proceedings of the National Academy of
Sciences of the United States of America, 98, 4534–4539. https​://doi.
org/10.1073/pnas.07103​4898
Ricotta, C. (2005). Through the jungle of biological diver‐
sity. Acta Biotheoretica, 53, 29–38. https​://doi.org/10.1007/
s10441-005-7001-6
Riera, R., Tuset, V. M., Betancur‐R, R., Lombarte, A., Marcos, C., & Pérez‐
Ruzafa, A. (2018). Modelling alpha‐diversities of coastal lagoon fish
assemblages from the Mediterranean Sea. Progress in Oceanography,
165, 100–109. https​://doi.org/10.1016/j.pocean.2018.05.003
Santini, L., Belmaker, J., Costello, M. J., Pereira, H. M., Rossberg, A. G.,
Schipper, A. M., … Rondinini, C. (2017). Assessing the suitability of di‐
versity metrics to detect biodiversity change. Biological Conservation,
213, 341–350. https​://doi.org/10.1016/j.biocon.2016.08.024
Schliep, E. M., Lany, N. K., Zarnetske, P. L., Schaeffer, R. N., Orians, C.
M., Orwig, D. A., & Preisser, E. L. (2018). Joint species distribution
modelling for spatio‐temporal occurrence and ordinal abundance
data. Global Ecology and Biogeography, 27, 142–155. https​://doi.
org/10.1111/geb.12666​
Schmitt, S., Pouteau, R., Justeau, D., de Boissieu, F., & Birnbaum, P.
(2017). ssdm: An r package to predict distribution of species rich‐
ness and composition based on stacked species distribution models.
Methods in Ecology and Evolution, 8, 1795–1803.
Singer, A., Johst, K., Banitz, T., Fowler, M. S., Groeneveld, J., Gutiérrez, A.
G., … Travis, J. M. J. (2016). Community dynamics under environmen‐
tal change: How can next generation mechanistic models improve
projections of species distributions? Ecological Modelling, 326, 63–74.
https​://doi.org/10.1016/j.ecolm​odel.2015.11.007
Skrondal, A., & Rabe‐Hesketh, S. (2004). Generalized latent variable mod‐
eling: Multilevel, longitudinal, and structural equation models. New
York, NY: Chapman and Hall/CRC.
Stockwell, D. R. B., & Peterson, A. T. (2002). Effects of sample size on ac‐
curacy of species distribution models. Ecological Modelling, 148, 1–13.
https​://doi.org/10.1016/S0304-3800(01)00388-X
Tessarolo, G., Rangel, T. F., Araújo, M. B., & Hortal, J. (2014). Uncertainty
associated with survey design in Species Distribution Models.
Diversity and Distributions, 20, 1258–1269. https​://doi.org/10.1111/
ddi.12236​
Thorson, J. T., & Barnett, L. A. K. (2017). Comparing estimates of abun‐
dance trends and distribution shifts using single‐ and multispecies
models of fishes and biogenic habitat. ICES Journal of Marine Science:
Journal Du Conseil, 74, 1311–1321. https​://doi.org/10.1093/icesj​ms/
fsw193
Thuiller, W., Münkemüller, T., Lavergne, S., Mouillot, D., Mouquet, N.,
Schiffers, K., & Gravel, D. (2013). A road map for integrating eco‐
evolutionary processes into biodiversity models. Ecology Letters, 16,
94–105. https​://doi.org/10.1111/ele.12104​
 |
1708       ZHANG et al.

Tuomisto, H. (2010). A diversity of beta diversities: Straightening

B I O S K E TC H
up a concept gone awry. Part 2. Quantifying beta diversity
and related phenomena. Ecography, 33, 23–45. https​://doi.
Chongliang Zhang is a marine fisheries scientist. His research in‐
org/10.1111/j.1600-0587.2009.06148.x
Urban, M. C., Bocedi, G., Hendry, A. P., Mihoub, J.‐B., Peer, G., Singer, A., terests are ecosystem modelling to support marine and fisheries
… Travis, J. M. J. (2016). Improving the forecast for biodiversity under management.
climate change. Science, 353, aad8466. https​://doi.org/10.1126/
scien​ce.aad8466 Author contributions: C.Z. conceived the ideas; C.Z. and Y.C. led
Warton, D. I., Blanchet, F. G., O'Hara, R. B., Ovaskainen, O., Taskinen, S., the analysis and writing, assisted by B.X., Y.X. and Y.R. All authors
Walker, S. C., & Hui, F. K. C. (2015). So many variables: Joint modeling contributed to the interpretation of analyses.
in community ecology. Trends in Ecology and Evolution, 30, 766–779.
https​://doi.org/10.1016/j.tree.2015.09.007
Whittaker, R. H. (1960). Vegetation of the Siskiyou Mountains, Oregon
S U P P O R T I N G I N FO R M AT I O N
and California. Ecological Monographs, 30, 280–338. https​://doi.
org/10.2307/1943563 Additional supporting information may be found online in the
Wisz, M. S., Pottier, J., Kissling, W. D., Pellissier, L., Lenoir, J., Damgaard,
Supporting Information section at the end of the article. 
C. F., … Svenning, J.‐C. (2013). The role of biotic interactions in shap‐
ing distributions and realised assemblages of species: Implications
for species distribution modelling. Biological Reviews, 88, 15–30.
https​://doi.org/10.1111/j.1469-185X.2012.00235.x How to cite this article: Zhang C, Chen Y, Xu B, Xue Y, Ren Y.
Zhang, C., & Chen, Y. (2015). Development of abundance indices for How to predict biodiversity in space? An evaluation of
Atlantic Cod and Cusk in the Coastal Gulf of Maine from their modelling approaches in marine ecosystems. Divers Distrib.
Bycatch in the Lobster Fishery. North American Journal of Fisheries
2019;25:1697–1708. https​://doi.org/10.1111/ddi.12970​
Management, 35, 708–719. https​://doi.org/10.1080/02755​
947.2015.1043413
Zhang, C., Chen, Y., Xu, B., Xue, Y., & Ren, Y. (2018). Comparing the pre‐
diction of joint species distribution models with respect to charac‐
teristics of sampling data. Ecography, 41, 1876–1887. https​://doi.
org/10.1111/ecog.03571​