Jukuri, open repository of the Natural Resources Institute Finland (Luke) 
   
 
   
All material supplied via Jukuri is protected by copyright and other intellectual property rights. Duplication 
or sale, in electronic or print form, of any part of the repository collections is prohibited. Making electronic 
or print copies of the material is permitted only for your own personal use or for educational purposes.  For 
other purposes, this article may be used in accordance with the publisher’s terms. There may be 
differences between this version and the publisher’s version. You are advised to cite the publisher’s 
version. 
 
This is an electronic reprint of the original article.  
This reprint may differ from the original in pagination and typographic detail. 
 
Author(s): Parvez Rana & Anne Tolvanen 
 
Title: Transferability of 34 red-listed peatland plant species models across boreal vegetation 
zone 
Year:  2021 
Version: Publisher’s version 
Copyright:    The author(s) 2021   
Rights:  CC BY  
Rights url: https://creativecommons.org/licenses/by/4.0/  
 
Please cite the original version: 
Parvez Rana & Anne Tolvanen. Transferability of 34 red-listed peatland plant species models across boreal 
vegetation zone. Ecological indicators 2021, 129, 107950. https://doi.org/10.1016/j.ecolind.2021.107950 
 
Ecological Indicators 129 (2021) 107950
Available online 8 July 2021
1470-160X/© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Transferability of 34 red-listed peatland plant species models across boreal 
vegetation zone 
Parvez Rana *, Anne Tolvanen 
Natural Resources Institute Finland (Luke), Paavo Havaksen tie 3, 90570 Oulu, Finland   
A R T I C L E  I N F O   
Keywords: 
Biodiversity 
Habitat modeling 
MaxEnt 
Peatlands 
Boreal forests 
A B S T R A C T   
Red-listed species are negatively affected by habitat degradation and fragmentation. They usually have small 
populations and may be affected by local extinctions which makes species distribution modeling (SDM) chal-
lenging. Although SDM has been used extensively for biodiversity protection and regional land-use planning, the 
transferability of SDM between regions is still at the developmental level. We show a first attempt to demonstrate 
model transferability for red-listed plant species over boreal regions. We modeled the distribution of 34 red-listed 
boreal peatland plant species at national and regional levels, using multiple streams of environmental data in 
Finland. The objectives were: (1) to evaluate how environmental characteristics explaining species distribution 
differ between three regions covering five vegetation zones (subarctic, northern, middle, southern boreal, and 
hemi-boreal vegetation zones); (2) to assess the performance of one national and three regional species distri-
bution models (SDM: northern, middle, and southern regions); and (3) to test whether the regional models can be 
transferred to other regions and discuss alternative methods to improve transferability. The maximum entropy 
(maxent) algorithm was employed to predict suitable habitats for the assessed species. An SDM performance was 
measured with the area under the receiver operative characteristics (AUC), true skill statistics (TSS), and the 
continuous Boyce index (CBI). Three conclusions are relevant. First, the environmental variables explaining 
species distribution differed significantly (p < 0.05) between the three regions. Second, the internal measure of 
accuracy measured as cross-validation of AUC, TSS, and the CBI was quite similar in both the national and 
regional models, which indicates that realistic species distribution maps could be generated from all models. 
Last, the external measure of accuracy (i.e. transferability) in the regional models was lower than the internal 
measure of accuracy, which indicates that a good regional model could not automatically ensure good perfor-
mance when transferred to another region. To improve the transferability of the regional models, we suggest the 
normalizing of environmental variable values. The data-driven evaluation of red-listed plant species provides an 
approach that can be used in biodiversity and nature conservation.   
1. Introduction 
The biodiversity conservation goal has gained increasing consider-
ation globally due to the significant loss of floral and faunal diversity 
(IPBES, 2018). The Convention on Biological Diversity (CBD) estimates 
that 30% of the world’s land needs to be protected in the next decade to 
halt the decline and extinction of global biodiversity (CBD, 2020). 
Degradation and fragmentation can negatively affect the distribution of 
red-listed species (Hanski, 2015). For example, in boreal Finland, a 
growing trend of the number of red-listed species (all taxa included) was 
evidenced, with an increase of 1,020 species between 2010 and 2019 
(Hyvarinen et al., 2019). Every ninth species was considered red-listed, 
the major risk being the degradation of suitable habitats caused by 
disturbance, construction, forest management activities, the reduction 
of old-growth forests, and climate change (Hyvarinen et al., 2019). To 
halt the biodiversity loss, international targets are being updated (CBD, 
2020) or have just been set (i.e. European Union (EU) Biodiversity 
Strategy for 2030, European Commission, 2020) to protect significant 
amounts of the land and the sea area, or to restore biodiverse areas with 
high ecosystem services potential (e.g. biomass). The identification of 
potential nature protection sites and prediction of suitable habitats for 
red-listed species are core elements of future biodiversity protection and 
ecosystem restoration. 
Peatlands (wetlands with peat soil) harbor a rich and unique range of 
* Corresponding author. 
E-mail address: parvez.rana@luke.fi (P. Rana).  
Contents lists available at ScienceDirect 
Ecological Indicators 
journal homepage: www.elsevier.com/locate/ecolind 
https://doi.org/10.1016/j.ecolind.2021.107950 
Received 5 May 2021; Received in revised form 29 June 2021; Accepted 2 July 2021   
Ecological Indicators 129 (2021) 107950
2
species and habitats worldwide (e.g. Yu, 2012). They are a highly space- 
effective reservoir of carbon and provide a wide set of other ecosystem 
services such as recreation, water management, timber, energy, and 
horticultural peat (e.g. Zedler and Kercher, 2005; Byg et al., 2017; 
Heikkinen et al., 2018). Globally, the area of peatland cover has been 
reduced to 10–20% (Cris et al., 2014), and 15% of peatlands have been 
drained, which is the main cause of change in peatland biodiversity 
(Harpenslager et al., 2015). Peatlands cover about 30% (approx. 10 
million hectares) of the land area (Turunen, 2008) of Finland. A large 
percentage of forestry drainage (more than 50%, Finnish Forest 
Research Institute, 2014), conversion to agricultural uses, and extrac-
tion of peat and minerals are the main causes of peatland habitat 
degradation (Hyvarinen et al., 2019). Using predictive models, it was 
estimated that the 15% ecosystem restoration target set in the previous 
CBD and EU Biodiversity Strategy to 2020 targets would increase the 
predicted distribution of 71% of red-listed peatland plant species, 
whereas the potentially occupied area of species would increase only by 
9% (Tolvanen et al., 2020). The models were developed for the aapa 
mire region that covers part of the northern and middle boreal regions 
(Saarimaa et al., 2019; Tolvanen et al., 2020). A new methodology is 
therefore needed that is transferable to the vegetation zones (i.e. sub-
arctic, northern, middle, southern, and hemi-boreal zones) where boreal 
red-listed peatland plant species are found. Because of the large longi-
tudinal dimension from the 60N hemi-boreal vegetation zone to the 
70N subarctic zone, the diversity of peatland characteristics is large in 
Finland (Seppa, 2002). However, the environmental characteristics (e.g. 
temperature, rainfall) and the diversity of peatland site types reflect the 
richness of peatland vegetation in all vegetation zones (Aapala et al., 
1998). 
Species distribution modeling (SDM, also known as habitat suit-
ability modeling) is widely used in biodiversity protection and regional 
land-use planning. The basic principle of SDM relies on correlating 
species occurrences with environmental variables, and employing geo-
statistical models or hypothetically inferred response curves (e.g. Araújo 
et al., 2019). Environmental data and recent progress in methods for 
SDM (e.g. Guisan and Thuiller, 2005; Zhang et al., 2018; Arenas-Castro 
et al., 2020; Li et al., 2020) provide opportunities to estimate species 
occurrences on a large spatial scale. Despite the wide use of SDMs, their 
usefulness as conservation tools has been doubted due to their tendency 
to have a larger internal measure of accuracy (also known as interpo-
lation) and a lower external measure of accuracy (also known as 
transferability or extrapolation) (Datta et al., 2020; Heikkinen et al., 
2012; Huang et al., 2018; Low et al., 2021; Manzoor et al., 2018; Qiao 
et al., 2019). These characteristics can constrain the reliable application 
of SDMs beyond surveyed regions. In addition, existing biodiversity 
models still struggle with data limitation, model transferability limita-
tion (i.e. the extrapolation or generalization of the models), and complex 
structure of biological systems (Araújo et al., 2019). Earlier studies 
concentrated only on a small area within one vegetation zone (Par-
viainen et al., 2009, 2008, 2013) or on Aapa mires that cover parts of 
northern and middle boreal zones in Finnish boreal peatlands (Saarimaa 
et al., 2019; Tolvanen et al., 2020). In addition, the generalization of the 
models over other areas or areas covering multiple vegetation zones was 
missing the previous studies. Model transferability into new areas or 
other time frames is one of the challenges of SDM (Datta et al., 2020; 
Low et al., 2021; Qiao et al., 2019). Using SDM in an extrapolative 
manner (transferability or external measure of accuracy) is a complex 
task compared to the simple filling of gaps in SDM within partly sampled 
areas (i.e. interpolative estimation or internal measure of estimation) 
(Heikkinen et al., 2012; Huang et al., 2018; Manzoor et al., 2018). The 
trained SDM may fail to represent distinct environmental characteristics 
in the unsurveyed areas. Some recent SDMs use deep learning algo-
rithms (e.g., convolutional neural networks) to better address the 
complexity of ecological niches (Deneu et al., 2021). 
SDMs for predicting red-listed plant species distribution are 
becoming increasingly important to assess their extinction risk and the 
degree of protection (Marcer et al., 2013; McCune, 2016; Parviainen 
et al., 2008, 2013; Saarimaa et al., 2019; Thapa et al., 2020). In addition, 
SDM-derived hotspot maps can be used in tandem with spatial conser-
vation prioritization to protected area network mapping (Marcer et al., 
2013; McCune, 2016; Parviainen et al., 2013). However, the scarce 
occurrence data continues to challenge the use of SDM for red-listed 
plant species (Marcer et al., 2013; McCune, 2016). In previous studies, 
researchers have focused on optimizing SDM and interpolation perfor-
mance of SDM (e.g. Bellamy et al., 2013; McCune, 2016; Arenas-Castro 
et al., 2020), but not on the extrapolation performance (Datta et al., 
2020; Heikkinen et al., 2012; Randin et al., 2006; Wenger and Olden, 
2012). Transferability assessment of SDM in spatio-temporal scale pro-
vides crucial information for red-listed plant species (Heikkinen et al., 
2012) as it helps to evaluate whether there is suitable habitat for the 
species outside their present distribution range. This information helps 
the future management and protection measures of the species. Trans-
ferability of an SDM may also be important in northern ecosystems 
where species distributions are predicted to change more rapidly than in 
regions further south. 
In this study, we modeled the distribution of 34 red-listed boreal 
peatland plant species at national and regional levels. Our objectives 
were (1) to evaluate how environmental characteristics explaining 
species distribution differ between three regions covering five vegeta-
tion zones (subarctic, and northern, middle, southern boreal, and hemi- 
boreal vegetation zones); (2) to assess the performance of the national 
and three regional SDM (northern, middle, and southern regions); and 
(3) to test the transferability of regional models to other regions and 
discuss alternative methods to improve the transferability. We utilized 
nationwide species occurrence data observed for the last 20 years and 
included four categories of environmental variables (i.e. climate, 
topography, geology, and habitat characteristics) as explanatory vari-
ables in the models. We extended the earlier models (Saarimaa et al., 
2019; Tolvanen et al., 2020) to cover the five vegetation zones in 
Finland in which the studied red-listed species have been encountered. 
We generalized the results over the regions using a maximum entropy 
(Maxent) algorithm. To our knowledge, this study was the first attempt 
to demonstrate model transferability for red-listed species distribution 
over boreal regions. 
2. Materials and methods 
2.1. Study area 
The study area covers the whole of Finland (60W, 70N) in northern 
Europe (Fig. 1). The climate is mild, influenced by the Gulf Stream. The 
annual average temperature is 6 C in the southwestern part of the 
country and  2 C in the northeastern part of the country. Across the 
country, precipitation varies between 400 mm and 700 mm. The study 
area is mostly flat terrain, with an area of 75% forest cover (Lier et al., 
2019). 
There are two types of peatland in Finland: Aapa mires, which are 
located mostly in the middle and northern boreal vegetation regions; 
and raised bogs, located mainly in the southern boreal region. Aapa 
mires are fens characterized by sedges and moss peat, whereas raised 
bogs are composed of Sphagnum peat and Eriophorum peat. 
The study area covered five vegetation zones (i.e. subarctic, and 
northern, middle, southern boreal, and hemi-boreal vegetation zones) 
(Ahti et al., 1968), which were merged into the northern region (sub-
arctic and northern boreal vegetation zones), middle region (middle 
boreal vegetation zone), and southern region (southern boreal and hemi- 
boreal vegetation zones) (Fig. 1) to increase the species occurrences in 
each region. The study area was segmented into a spatial grid of 500 m 
by 500 m (25 ha) squares, and squares with more than 10% peatland 
cover were retained in the analysis. As a result, we had a total of 
2,941,262 spatial grid squares (735,316 km2) for analysis. 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
3
2.2. Red-listed plant species data 
Peatlands are the primary habitat of 280 red-listed species, which is 
4.5% of the total red-listed species in Finland (Hyvarinen et al., 2019). 
The national database was utilized to extract the presence-only occur-
rence records of red-listed peatland plant species (Hyvarinen et al., 
2019). The term red-listed includes the IUCN categories critically en-
dangered (CR), endangered (EN), vulnerable (VU), and near threatened 
(NT). We utilized the following standards for data extraction from the 
database: (i) the occurrences of those vascular plants and bryophytes 
which occur primarily on undrained mires; (ii) the newest occurrence 
records (i.e. from 2000 or later); (iii) only observations with accuracy 
better than 100 m; (iv) species with at least ten or more records of 
occurrence. Species absences are not included in the database. Finally, 
we had a total of 8,846 occurrence records of 34 red-listed plant species, 
which were used for model calibration and validation (Appendix 1). 
The red-listed plant species (n ˆ 34) were subdivided into four 
species groups according to their principal habitats (see also Saarimaa 
et al., 2019): spruce swamp forest species (SSF); mesotrophic fen spe-
cies; rich fen species; and calcareous fen species. Spruce swamp forests 
are wooded minerotrophic nutrient-rich peatlands, characterized by 
Norway spruce (Picea abies), sparse deciduous trees, and abundant 
deadwood (Simila et al., 2014). Mesotrophic fens are characterized by 
thick peat layers formed from Sphagnum moss and other bryophytes, and 
Fig. 1. Map of the study area, with red-listed plant species occurrence records. The northern region includes northern boreal and subarctic vegetation regions, the 
middle region includes the middle boreal vegetation region, and the southern region includes southern boreal and hemi-boreal vegetation regions. (For interpretation 
of the references to colour in this figure legend, the reader is referred to the web version of this article.) 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
4
by sedges and herbaceous plants in the field layer. The main attributes of 
rich fens are large species diversity, neutral or mildly acidic soil, and an 
open or sparsely wooded landscape (Saarimaa et al., 2019). A large 
proportion of calcareous rock in the soil and high mineral concentration 
in groundwater are the main requirements for a calcareous fens habitat 
(Simila et al., 2014). The total number of species in each species group 
was 8, 5, 9, and 12 respectively for SSF, mesotrophic fen, rich fen, and 
calcareous fen species. The northern region had a total of 14 species, the 
middle region 21 species, and the southern region 13 species. 
2.3. Environmental data 
A total of 16 environmental variables within four categories (i.e. 
climate (n ˆ 2), topography (n ˆ 1), geology (n ˆ 1), and habitat (n ˆ
12)) were employed in the model building (Appendix 2). We ensured 
using Spearman’s rank correlation (pair-wise absolute correlation cutoff 
ˆ 0.70) that multicollinearity among the environmental variables was 
absent (e.g. Jimenez-Valverde and Lobo, 2007; McCune, 2016). In the 
climate category, growing degree days (GDD) and mean water balance 
were calculated for 1990–2013 from the climate data of the Finnish 
Meteorological Institute (Pirinen et al., 2012). Despite that the most 
recent data (2014–2019) were not available at the same resolution, the 
23-year time period (1990–2013) covered a considerable amount of the 
years when species observations had been made. Hence the data could 
capture the GDD and mean water balance conditions of the species quite 
well. The GDD reflecting the extent of the growing season was calculated 
according to Equation (1). 
GDD ˆ
X
maxf0; …Tm  Tt†g (1)  
where Tm is the mean daily temperature, and Tt is a threshold temper-
ature (5 C). Precipitation alone is insufficient to indicate the moisture 
available for plant growth; water balance was therefore calculated as the 
monthly difference between precipitation and potential evapotranspi-
ration. The monthly water balance was summed per annum. The po-
tential evapotranspiration (PET) was calculated according to Equation 
(2) (Skov and Svenning, 2004). 
PET ˆ
58:93 T…above0 C†
12 (2) 
In the topography category, the topographic wetness index (TWI) 
(Beven and Kirkby, 1979) indicates the local relative differences in 
moisture conditions and was calculated using Equation (3). 
TWI ˆ ln
 α
tanβ

; (3)  
where α is the local upslope area draining through a certain point per 
unit contour length, and tanβ is the local slope in radians. A high TWI 
value indicates a wet area. 
In the geology category, the proportion of calcareous rock in each 
grid square was assessed by utilizing a digital map of the Quaternary 
deposit and pre-Quaternary rock database. 
In the habitat category, we employed 12 environmental variables, 
which provide information on tree species’ mean volume, site type, 
drainage status, and biodiversity index (BI) (Appendix 2). We derived 
eight variables from the multi-source national forest inventory (MS-NFI) 
data: the mean volume (m3/ha) of pine (Pinus sylvestries), spruce (Picea 
abies), birch (Betula pendula and B. pubescens), and other broadleaved 
trees, and the proportion of herb-rich Vaccinium myrtillus, Vaccinium 
vitis-idaea, and Cladina forest site types in each grid square (%). The 
proportion of undrained, drained, and open peatlands in each grid 
square was calculated using a drainage map by the Finnish Environment 
Institute (SYKE, 2009). The drainage map is based on the topographic 
database of the Finnish National Land Survey. Last but not least, we 
assessed a BI value developed for forests (Mikkonen et al., 2018) for each 
grid square. The BI value is normalized from 0 to 1, with high values 
indicating high biodiversity. The index is formed from three main ele-
ments: 1: local decaying wood potential, number of large trees, and tree 
species diversity; 2: penalties for forestry operations (e.g. logging, clear- 
cutting); and 3: connectivity, based on ecological similarity, distance, 
and quality between forest stands. For a detailed description, see Mik-
konen et al. (2018). 
2.4. Statistical testing 
Since our data were not normally distributed (confirmed by the 
Kolmogorov-Smirnov test, Marsaglia et al., 2003), we used non- 
parametric Wilcoxon signed rank to assess whether the environmental 
variable values differed between the three regions (Hollander and 
Wolfe, 1973). Testing was done in a pairwise manner: northern vs. 
middle region; northern vs. southern region; middle vs. southern region. 
We used Wilcoxon signed rank, which uses signed ranks and is an 
appropriate test in our study. 
The Bhattacharyya distance was used as a measure of divergence 
(Davis, 1996; Fukunaga, 2013) to realize the necessity of data normal-
ization for model transferability (Huang et al., 2018; Qiao et al., 2019). 
The advantage of the Bhattacharyya distance separability criterion is the 
fact that it is a simple criterion that can be easily implemented. The 
Bhattacharyya distance can be defined as follows: Let i and j be the two 
variable distributions that are usually modeled as a Gaussian distribu-
tion with mean µ and covariance 
P
. The Bhattacharyya distance be-
tween two probability distributions Pi and Pj denoted by Bij is defined by 
Bij ˆ
1
8

μi  μj

T
…
P
i ‡
P
j
2 †
 1 μi  μj


‡
1
2 ln
0
B
@

Σi ‡ Σj



jΣij

Σj


q
1
C
A; (4)  
where μi and μj refer to the mean of the distributions, 
P
i and 
P
j refers to 
the covariance matrix of the distributions. The Bhattacharyya distance 
of environmental variable distributions between two regions (e.g. 
northern vs. middle region) will be zero if there is a complete overlap 
(Davis, 1996; Fukunaga, 2013). 
2.5. Model calibration and validation 
The overall workflow, from the initial environmental input files and 
species occurrences to the delineated hotspot maps, is presented in 
Fig. 2. 
We employed a maxent algorithm to predict the habitat suitability 
for the red-listed plant species. Maxent generates a continuous predic-
tion of occurrences, ranging from 0 (least suitable) to 1 (best suitable) (e. 
g. Elith et al., 2011; Elith and Leathwick, 2009; Low et al., 2021; Merow 
et al., 2013; Phillips et al., 2006; Phillips and Dubik, 2008; Saarimaa 
et al., 2019). It has been shown effective in predicting the distributions 
of peatland plant species (McCune, 2016; Oke and Hager, 2017; Par-
viainen et al., 2013; Saarimaa et al., 2019; Tolvanen et al., 2020). We 
applied the default parameters, i.e. regularization multiplier ˆ 1, auto- 
features, maximum iterations ˆ 500, and convergence threshold ˆ 10 5, 
to generate reasonably simple models that did not over-fit the calibra-
tion data (Kramer-Schadt et al., 2013; Phillips and Dubik, 2008; Virzi 
et al., 2017). The logistic output was used to compare the prediction 
maps between species (Li et al., 2020; Saarimaa et al., 2019). We 
employed the same maxent setting and the same environmental vari-
ables for all species so that we could aggregate models for several 
species. 
Since we had presence-only data, we used pseudo-absences in the 
models to help control the sampling bias, and avoid erroneous or false 
absences (Bellamy et al., 2013; Brown et al., 2017). Following the pre-
vious literature (e.g. Mateo et al., 2010; Barbet-Massin et al., 2012; 
Connor et al., 2018), we generated 10,000 pseudo-absence points, and 
we gave equal weight to the presences and pseudo-absence records 
(prevalence of 0.5) (Elith et al., 2011; Jimenez-Valverde and Lobo, 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
5
2007; Liu et al., 2016; Tehrani et al., 2020). We sampled the pseudo- 
absence points with a buffered minimum-convex polygon (buffer 
radius ˆ 100 km, labeled as bias grid) based on known occurrences for 
each species (Brown et al., 2017). Bias grids regulate the pseudo-absence 
point’s allocation and sampling density. 
We assessed the SDMs at the national and regional levels, i.e in the 
northern, middle and southern regions (Fig. 2). In the national model, 
we postulated that we could build species-level SDMs that covered all 
occurrences of the respective species in all three regions. We calculated 
k-fold cross-validation (k-fold CV) accuracy metrics (see Section 2.6 for 
details) and employed k ˆ 5 with a random partition of 80% training and 
20% validation (Saarimaa et al., 2019; Tehrani et al., 2020; Araújo et al., 
2005). 
In the regional models, we postulated that we could build species- 
level SDMs for each region and transfer them to other regions that 
were not covered by the training data. For example, we built a model 
using all occurrence records of a species (e.g. Carex appropinquata) found 
in the northern region and reported the internal measure of accuracy 
metrics. We then validated (i.e. transferability) it on the other region (i. 
e. middle and southern regions) and reported the external measure of 
accuracy metrics. Similarly, we repeated the above calibration and 
validation for the middle and southern regions. 
2.6. Assessing model performance and key environmental variable 
The area under the receiver operative characteristics (AUC) offers 
widely used statistics for assessing maxent model performance (Marcer 
et al., 2013; Parviainen et al., 2009, 2008, 2013; Saarimaa et al., 2019; 
Syfert et al., 2014; Zhang et al., 2018), although it has recently been 
criticized (Fernandes et al., 2019; Manzoor et al., 2018). True skill sta-
tistics (TSS, Allouche et al., 2006; Fernandes et al., 2019; Li et al., 2020; 
Tehrani et al., 2020), and the continuous Boyce index (CBI, Hirzel et al., 
2006; Progin, 2018) were therefore considered the most reliable statis-
tics for evaluating presence-only data (Zurell et al., 2020). CBI is also an 
appropriate metric to validate model transferability (Cianfrani et al., 
2010; Manzoor et al., 2018). Zurell et al. (2020), however, reported a 
standard protocol for reporting SDM performance where AUC, TSS and 
CBI were advised for evaluating presence-only data and this is why we 
decided to use these statistical metrics. 
The AUC, TSS, and CBI values were calculated for each species and 
species group (see Section 2.2 for species group description). AUC value 
ranges between 0 and 1 values < 0.7 were considered fair, 0.7–0.8 good, 
0.8–0.9 very good, and >0.9 excellent (Manzoor et al., 2018; Swets, 
1988). The range of TSS is  1 to ‡ 1, where values close to ‡ 1 indicate 
that the model can distinguish between false negative and false positive 
occurrences. The CBI avoids using a confusion matrix and is threshold- 
independent (Hirzel et al., 2006; Low et al., 2021). The CBI ranges be-
tween  1 and ‡ 1, where values close to ‡ 1 indicate that the predicted 
model distribution is consistent with the distribution of presences in the 
evaluation dataset (Boyce et al., 2002; Hirzel et al., 2006). 
We employed maxent permutation importance analysis to identify 
key environmental variables for each species, and we then summarized 
the permutation percentage values for each species group (Table 3). In 
assessing the contribution of each variable, maxent permutation 
importance analysis performs better than percentage contribution 
analysis, because permutation importance is path (algorithm) -inde-
pendent (e.g. Saarimaa et al., 2019), whereas the percentage contribu-
tion is based on the order of the variable in a model. The contribution of 
each variable relies on randomly permuting the values of that variable 
among the training points (both presence and pseudo-absence), and the 
final values are normalized to percentages (Connor et al., 2018; Li et al., 
2020; Saarimaa et al., 2019). 
2.7. Hotspot maps 
To produce hotspot maps for the species groups and for all species, 
we utilized the predicted distribution map of the national model, 
because the national model covered the entire study area (Figs. 6 and 7). 
In addition, we produced a hotspots map using finer grain data of 25 ha 
(0.25 km2), because a criticism has been that coarse-grain data (e.g. 10 
km2) may overestimate the hotspots value (Di Marco et al., 2017; 
McKerrow et al., 2018). Maxent produced a continuous probability map, 
Fig. 2. Overall flowchart of the model building and validation.  
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
6
which was reclassified into a binary map for simplicity, presenting un-
suitable (0) and suitable (1) areas. Threshold values for maximum 
sensitivity and specificity (Max SS) were used for the reclassification. 
The Max SS threshold sets the limit at which the sum of sensitivity (true 
positive rate) and specificity (true negative rate) is highest (Liu et al., 
2013). The Max SS also reduces the mean error rate and minimizes the 
risk of selecting inappropriate areas by identifying only those areas with 
the maximum suitability (Saarimaa et al., 2019). 
We stacked the individual species-level suitability values for the 
species groups and created a hotspot map for each species group. The top 
5% of grid squares ranked by stacked suitability value were presented in 
the map (e.g. Williams et al., 1996; Orme et al., 2005). To produce a 
hotspot map for all red-listed species (n ˆ 34), the summing of suit-
ability values was done across all species, and 5% of the grid squares 
ranked by stacked suitability value were presented. 
2.8. Data processing and analysis environment 
Data processing, mapping and statistical analysis were done using R 
statistical programming (libraries, i.e. raster, dismo, SDMtune, ecospat, 
sf, rJava) (R Core Team, 2020), Python programming (libraries, i.e. 
arcpy, pandas, numpy), and ArcGIS version 10.6.1 (ESRI, 2018). 
3. Results 
3.1. Environmental variables in the three regions 
The distribution of 13 out of 16 environmental variables was sta-
tistically different (p < 0.05) between all three regions (Fig. 3, Fig. 4, 
Table 1, Appendix 3). For example, temperatures (GDD) were coldest in 
the northern region and warmest in the southern region (Appendix 3), 
which was confirmed by the large Bhattacharyya distance (3.34) be-
tween the northern and southern regions (Table 1) and by the proba-
bility density function where northern and southern regions were not 
overlapping (Fig. 3). TWI was less varied in the three modeling regions 
which was visible in the Bhattacharyya distance (Table 1) and proba-
bility density function (Fig. 3). The proportion of calcareous rock was 
largest in the northern region, whereas the middle vs. the southern re-
gion showed no difference (Bhattacharyya distance 0.0). Average pine, 
spruce, and birch tree volumes (m3/ha) were largest in the southern 
region (Appendix 3), which was confirmed by the large Bhattacharyya 
distance between the northern and southern regions (Table 1). Tree 
volumes distribution differences were also visible in the probability 
density function, for instance, the northern region had a spike at zero 
percent of their pine tree volume distribution (Fig. 3). The proportion of 
the Vaccinium myrtillus site type was largest in the southern region 
(Appendix 3), and the Bhattacharyya distance was largest (0.37) be-
tween the middle and southern regions (Table 1). The percentage of 
undrained peatlands was highest in the northern region and lowest in 
Fig. 3. Growing degree days, topographic wetness index, and mean volume of pine in the three regions. Wilcoxon signed-rank test confirmed that the distributions 
were statistically significantly different (p < 0.05) in the three regions for model building. 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
7
the southern region (Appendix 3), which was confirmed by the large 
Bhattacharyya distance (0.43) between the northern and southern re-
gions (Table 1) and by the probability density function where northern 
and southern regions were less overlapping (Fig. 4). 
3.2. National models 
The national model performed well for all 34 plant species, which 
was shown by all three accuracy metrics. The mean AUCs were excellent 
(average 0.97, range 0.89–1) and very good (average 0.88, range 0.47–1) 
for training and cross-validation respectively (Fig. 5, Appendix 4). Only 
for two species (Plagiothecium latebricola, Sphagnum subnitens) were the 
mean cross-validation AUC values <0.70 (fair). The mean TSSs values 
were 0.87 (range 0.62–1) and 0.76 (range 0.42–0.99) for training and 
cross-validation respectively (Appendix 4, Fig. 5). The mean CBI was 
0.91 (range 0.74–1). 
Concerning the species groups, the highest mean (SD) AUC value of 
0.91 (0.09) at validation was recorded for the calcareous species, 
whereas the lowest value of 0.81 (0.15) was recorded for SSF species 
(Table 2). The TSS value showed a similar trend with AUC values. The 
CBI was highest for the mesotrophic fen species and lowest for the SSF 
species. 
GDD, the proportion of undrained peatlands, and the BI were the 
three most important variables in the national models (Table 3). The 
GDD had the most relative importance (%) for the mesotrophic fen 
species (52.3%), for SSF (30%), and for calcareous species (21.1%), 
whereas the proportion of undrained peatland areas had the most rela-
tive importance (36.2%) for the rich fen species. The BI had the most 
relative importance (19.4%) for the SSF species. 
Fig. 4. Proportion of Vaccinium vitis-idaea site type, proportion of undrained peatlands, and biodiversity index in the three regions. Wilcoxon signed-rank test 
confirmed that the distributions were statistically significantly different (p < 0.05) in the three regions for model building. 
Table 1 
Bhattacharyya distances of environmental variables between the three regions. 
The distance between the two regions is zero if there is complete overlap.  
Environmental 
variable 
Northern vs. 
middle region 
Northern vs. 
southern region 
Middle vs. 
southern region 
GDD 1.36  3.34 0.84 
WAB 0.04  0.07 0.09 
TWI 0.00  0.02 0.01 
Rock 0.04  0.02 0.00 
Pine 0.09  0.12 0.01 
Spruce 0.20  0.53 0.13 
Birch 0.12  0.25 0.03 
Broadleaf other 0.27  0.65 0.16 
Site type 1 0.01  0.02 0.06 
Site type 2 0.06  0.17 0.37 
Site type 3 0.00  0.02 0.04 
Site type 4 0.08  0.02 0.04 
Undrained 0.14  0.43 0.10 
Drained 0.16  0.06 0.04 
Open 0.54  0.43 0.02 
BI 0.01  0.05 0.05 
Abbreviations are explained in Appendix 3. 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
8
3.3. Regional models 
The regional models performed well, which was shown by all three- 
accuracy metrics (Fig. 5, Appendix 5, 7, 9). For example, all the SDMs 
had a cross-validation AUC value that was higher than 0.84, which in-
dicates the consistent excellent performance of regional models. Similar 
results were observed with the CBI values. The median value of cross- 
validated AUC and CBI values were close to each other in three 
regional models (Fig. 5). However, the hinges (the 25th and 75th per-
centiles) and whiskers (upper and lower whisker extends) of cross- 
validated TSS value were higher in the middle region compared to the 
northern and southern regions (Fig. 5). 
The external measure of accuracy (i.e. transferability) was lower 
than the internal measure (Fig. 6), which indicates that regional models 
provided more accurate prediction in internal estimation (filing the gaps 
within the region) than external estimation (predictions outside the 
region). This suggests that the transferability of regional models into 
new regions is challenging. The highest mean  SD transferability of 
AUC, TSS, and CBI was found in the northern region (0.76  0.01, 0.47 
 0.02, 0.69  0.16), whereas the lowest transferability AUC was found 
in the middle region (0.67  0.03, 0.38  0.03, 0.50  0.18) (Appendix 
6, 8, 10). Furthermore, 2 (out of 14, northern region), 10 (out of 21, 
middle region), and 6 (out of 13, southern region) species SDMs had an 
external validation AUC value of <0.7 (fair). In the northern region, the 
most important variables were GDD, the proportion of undrained peat-
land, and drained peatlands respectively (Appendix 11). GDD, the pro-
portion of undrained peatland, and BI were the three most important 
variables respectively in the middle and southern regions (Appendix 12, 
13). 
3.4. Hotspot maps 
Suitable habitats (i.e. based on the stacked suitability value) for the 
calcareous species group were mostly predicted in the northern (65% of 
all suitable grid squares of calcareous species) and middle (28%) regions 
(Fig. 7), whereas suitable habitats for mesotrophic fen species were 
predicted especially in the middle region (82% of suitable grid squares) 
(Fig. 7). The rich fen species had very suitable habitats in both the 
northern (54%) and middle regions (42%) (Fig. 7), whereas it was 
predicted that the majority (84%) of suitable habitats for SSF species 
were in the southern region (Fig. 7). The all-species hotspots were 
scattered throughout the country, although they were mostly located in 
Fig. 5. Internal measure of accuracy (i.e. cross-validation AUC, cross-validation TSS, and CBI) for validation data in national and regional models (i.e. northern, 
middle and southern regions). Boxplots visualize five summary statistics (the median, two hinges (the 25th and 75th percentiles) and two whiskers: the upper whisker 
extends from the hinge to the largest value; the lower whisker extends from the hinge to the smallest value). 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
9
the middle (48%) and northern regions (36%) (Fig. 8). 
4. Discussion 
This study shows that although the national and regional SDM per-
formed very well, the transferability of the regional models to other 
regions was not as good. The reason appears to be the very different 
environmental characteristics between the regions, which limits the 
distributions of individual species. In addition, the statistical challenges 
distressed the transferability of the regional model as it forces the al-
gorithm to make predictions for outside of training environmental 
conditions, which may often be erroneous. To our knowledge, this was 
the first study to test whether the SDM of red-listed species could be 
transferred across boreal regions. We discuss the benefits, limitations, 
and implications of our method below. 
4.1. Benefits and limitations of SDM 
Our model was capable of producing a promising accuracy, and it 
was quite stable at both training and cross-validation stages, indicating 
that realistic species distribution maps can be generated from both na-
tional and regional modeling. Saarimaa et al. (2019) studied threatened 
plant species in aapa mires in Finland, recording a mean cross-validation 
AUC of 86%, which is close to our study (88%). Besides, Parviainen et al. 
(2013) utilized Landsat images to estimate the SDM of 28 red-listed 
plant species in northeastern Finland, mentioning an accuracy of 87%. 
However, it must be noted that apart from a larger study area covering 
the whole country, we used a different database from earlier studies 
(Parviainen et al., 2009, 2008, 2013; Saarimaa et al., 2019; Tolvanen 
et al., 2020) due to the newest red-list assessment available for this study 
(Hyvarinen et al., 2019). 
SDM can provide a useful means to predict species occurrences in 
unsurveyed areas (e.g. Datta et al., 2020; Heikkinen et al., 2012; Huang 
et al., 2018; Regos et al., 2019). However, the transferability of regional 
models is the key to a successful application of SDM, required in 
ecological and biodiversity conservation (Datta et al., 2020; Duque-Lazo 
et al., 2016; Huang et al., 2018; Qiao et al., 2019; Regos et al., 2019; 
Torres et al., 2015). However, the trained SDM from an area may not 
represent the distinct characteristics well (e.g. habitat structure and 
climate condition) in the unsurveyed areas (Datta et al., 2020; Huang 
et al., 2018; Low et al., 2021; Manzoor et al., 2018; Torres et al., 2015). 
The transferability of a regional model in our study performed similarly 
to each other. However, the external measure of accuracy (i.e. trans-
ferability) was lower than the internal measure of accuracy in all three 
regions. A good regional model thus does not automatically ensure 
equally good performance when they are transferred to a new case study 
area. 
Heikkinen et al. (2012) studied the model transferability of vascular 
Fig. 6. Internal (cross-validation) and external (transferability) measures of accuracy (i.e. AUC, TSS, and CBI) of regional models (i.e. northern, middle and southern 
regions). Boxplots visualizes five summary statistics (the median, two hinges (the 25th and 75th percentiles) and two whiskers: the upper whisker extends from the 
hinge to the largest value; the lower whisker extends from the hinge to the smallest value). 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
10
plants in Finland, recording a mean transferability AUC value of 0.70, 
which resembled our estimation accuracy (0.73). However, our study 
modeling design is distinct from Heikkinen et al. (2012) in that we 
assessed the transferability of each regional model based on vegetation 
zone, whereas Heikkinen et al. (2012) combined occurrence records 
from all the vegetation zones. By mingling our study findings and the 
results of Heikkinen et al. (2012), we can see that the model trans-
ferability of plant species is a challenge, irrespective of modeling design. 
The transferability of regional models provides several benefits, 
which have also been observed in other studies (e.g. Randin et al., 2006; 
Heikkinen et al., 2012; Wenger and Olden, 2012; Datta et al., 2020). The 
temporal scale can be extended by generating new species distribution 
maps and hotspot maps when new environmental data is available over 
the same area, and the spatial scale can be expanded by applying the 
models to new areas. Assessing at a temporal scale would be a 
Fig. 7. Hotspots for each species group were based on the national model. The top 5% of grid squares ranked by stacked suitability value were presented in the map.  
Table 2 
Measure of accuracy for training and validation data for each species group 
summarized from Appendix 4, national model.  
Species group AUC 
training 
mean 
SD 
AUC 
validation 
mean  SD 
TSS 
training 
mean 
SD 
TSS 
validation 
mean  SD 
Internal 
CBI 
mean 
SD 
Calcareous 0.97 
0.03 
0.91 
0.09 
0.88 
0.12 
0.81 
0.12 
0.91 
0.05 
Mesotrophic fen 0.96 
0.03 
0.90 
0.06 
0.80 
0.10 
0.73 
0.13 
0.95 
0.03 
Rich fen 0.96 
0.03 
0.88 
0.11 
0.83 
0.10 
0.73 
0.12 
0.94 
0.04 
SSF 0.98 
0.01 
0.82 
0.15 
0.92 
0.08 
0.75 
0.15 
0.87 
0.07  
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
11
motivating methodological link for future investigation. 
To improve the transferability of the regional model, one possibility 
is to normalize the environmental variable values, which may differ 
considerably between regions. Non-linear response functions between 
species and environmental data may affect the transferability (Heikki-
nen et al., 2012). Possible normalizing methods could be histogram 
matching, median-based adjustment, and regression-based adjustment. 
Histogram matching is a widely used technique for the normalization of 
variable values employing the cumulative probability function (Gon-
zalez and Woods, 2017). For median-based adjustment, the median 
value per environmental variable is used to normalize the value between 
the reference area variable value (training dataset) and the dependent 
area variable value (validation dataset) (Kappal, 2019; Valikangas et al., 
2018). For regression-based adjustment, the variable normalization is 
based on simple linear regression, using the median value of variable per 
plant species. We suggest that future studies could test possible variable 
normalization methods that might improve the transferability of the 
regional model. 
The appropriate selection of pseudo-absence points is a crucial 
parameter in SDM, because it greatly influences the model’s accuracy 
and uncertainty (Alvarez-martinez et al., 2015; Benito et al., 2017; 
Hengl et al., 2009; Iturbide et al., 2015; Li et al., 2011; Norris et al., 
2011; Ranc et al., 2017; Senay et al., 2013; Wisz and Guisan, 2009). In 
this study, we randomly sampled the pseudo-absence points with a 
buffered minimum-convex polygon (bias grid), based on known occur-
rences for each species separately with a prevalence of 0.5. The bias grid 
helps to avert the sampling of habitat located outside a species’ known 
occurrences (Bellamy et al., 2013; Brown et al., 2017; Lentini et al., 
2018; Martin et al., 2020). However, to improve the model performance, 
alternative pseudo-absence point selection methods, e.g. random sam-
pling with environmental constraint (Iturbide et al., 2015; Lentini et al., 
2018), random sampling with a combination of geographical and 
environmental constraints (Iturbide et al., 2015; Senay et al., 2013), 
target group sampling (Andrew and Fox, 2020; Ranc et al., 2017), 
random sampling with k-means clustering (Iturbide et al., 2015; Senay 
et al., 2013), and random sampling with a self-organizing map (Koho-
nen, 2001), could be tested in future studies. 
SDM was also affected by other parameters, including threshold se-
lection during model development. An incorrect threshold can lead to an 
inferior SDM, with a false identification of suitable areas and hotspot 
maps (McKerrow et al., 2018; Trotta-Moreu and Lobo, 2010). However, 
threshold selection for each species during model building is neither an 
easy nor a straightforward task. In our study, we employed the 
maximum sensitivity and specificity threshold, which is quite practical, 
because it minimizes the risk of over prediction (Liu et al., 2013, 2016). 
Table 3 
Variable contribution (%) summarized for each species group in the national 
model.  
Geospatial 
variable 
SSF % Mesotrophic fen 
% 
Rich fen 
% 
Calcareous 
% 
Mean 
% 
GDD 33.01 52.26 21.24 21.1 31.90 
WAB 0.90 4.83 3.60 2.43 2.94 
TWI 1.65 0.39 0.80 4.30 1.79 
Rock 0.81 0.33 2.93 12.48 4.14 
Pine 7.43 0.57 1.55 4.44 3.50 
Spruce 1.42 1.35 2.25 1.07 1.52 
Birch 1.49 1.19 1.24 0.88 1.20 
Broadleaf 
other 
4.51 0.93 1.07 2.06 2.14 
Site type 1 1.18 0.96 2.29 6.54 2.74 
Site type 2 1.60 2.65 11.41 4.92 5.15 
Site type 3 8.03 1.42 2.86 3.15 3.87 
Site type 4 4.87 4.31 4.74 12.52 6.61 
Undrained 5.84 25.36 36.21 11.27 19.67 
Drained 7.61 2.08 2.30 1.79 3.45 
Open 0.25 0.06 0.07 0.17 0.14 
BI 19.38 1.31 5.44 10.88 9.25 
Total 100 100 100 100 100 
*Three most important variables for each species group and on average are in 
bold. Abbreviations are explained in Appendix 3. 
Fig. 8. A hotspot map of all red-listed plant species (n ˆ 34). The top 5% of grid squares ranked by stacked suitability value were presented in the map. (For 
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
12
Threshold selection can also be affected by the use of presence-only data, 
and even a small margin of the threshold can lead to poorer distribution 
maps (Jimenez-Valverde and Lobo, 2007; Liu et al., 2013, 2016). 
We used maxent algorithm for SDM which was extensively used in 
the literature (e.g. Elith et al., 2011; Elith and Leathwick, 2009; Low 
et al., 2021; Merow et al., 2013; Phillips et al., 2006; Phillips and Dubik, 
2008; Saarimaa et al., 2019). However, other algorithms for assessing 
the transferability of SDM can also be tested in the future, including deep 
learning algorithm, i.e., convolutional neural networks (CNN), radial 
basis function networks (RBFN), multilayer perceptrons (MP), long- 
short term memory (LSTM), Kohonen network or self-organizing map 
(SOM) (Botella et al., 2018; Deneu et al., 2021; Reisenbüchler et al., 
2021; Shamshirband et al., 2019; Zhou, 2020). The possible benefit of 
using a deep learning algorithm is to tackle the statistical challenges 
which concerned the transferability of the regional model. 
SDM was adopted to evaluate how environmental characteristics 
explaining species distribution differ between three regions in our study. 
However, the use of process-based model or biophysical model (Kearney 
and Porter, 2009), and geographically weighted regression (Paez et al., 
2011) might be a good alternative for future studies. Higgins et al. 
(2020) compared that process-based model may be more appropriate for 
making projection outside of the training data whereas, SDM has su-
perior ability to describe the data in the training area. 
4.2. Environmental variables influencing the distribution of the species 
groups 
Our models show that most hotspots are located in the middle and 
northern regions of the country. The situation is partly connected with 
the greater land-use pressure and low protection status of southern 
Finland, where only 5% of forest lands (including peatlands) are pro-
tected, compared to 80% in the north (LUKE, 2019). Furthermore, the 
unprotected old-growth forest area is <2% in the total forest area in the 
southern region (Tikkanen et al., 2006). Nevertheless, the greater 
coverage of hotspot areas in the middle and northern regions also re-
flects the different habitat requirements between the species groups and 
the availability of the preferred habitats. The northern region was most 
suitable for rich fen and calcareous species, whereas the southern region 
was dominated by SSF species. For example, calcareous habitats occur 
naturally in the northern part of the country. The middle region of the 
country was especially suitable for mesotrophic fen species. 
Our results support previous observations on the importance of GDD 
and undrained peatland variables for red-listed species in the aapa mire 
region (Parviainen et al., 2013; Saarimaa et al., 2019), confirming that 
these variables also explain their distribution in a larger geographical 
area. Although the choice of variables in the national and regional 
models was similar, the contribution of each variable varied between the 
models. GDD seems the most useful variable, because it provides crucial 
information about the plants’ growth conditions. The mean GDD 
differed between habitat groups, which indicates the specific GDD 
preferences in these groups. 
Undrained peatlands are characterized by high TWI, diverse soil 
fertility, diverse micro-topography, and high peat layers (Saarimaa 
et al., 2019; Sallinen et al., 2019). These characteristics favor rich fen 
and mesotrophic fen species. Undrained mires, and hence rich fen and 
mesotrophic fen species, are mostly located in the northern and middle 
regions of Finland, whereas drained peatlands are located in the 
southern region due to the denser human population, and more inten-
sive forest management and agriculture. In the southern region, peat-
lands are also smaller (e.g. < 2 ha). They are confined to separate basins, 
which makes them more exposed to disturbance than undrained mires 
(Sallinen et al., 2019). Since rich fen and mesotrophic fen species require 
undrained conditions, their habitat suitability decreases after drainage 
due to the drawdown of the water table. The choice of BI in the top list 
was unsurprising, because the BI includes forest-related variables, e.g. 
decaying wood potential, forest connectivity, and connectivity with the 
key woodland habitat, which are either directly or indirectly linked with 
the distribution of red-listed species, especially SSF. 
We used four categories (climate, topography, geology, and habitat) 
of environmental variables in this study, as our studied 34 red-listed 
plant species are well correlated with the 16 environmental variables 
in these four categories (e.g. Parviainen et al., 2008, 2009, 2013; Saar-
imaa et al., 2019; Tolvanen et al., 2020). New studies shown that snow 
persistence, freezing degree days, thawing degree days, and permafrost 
occurrence may have an impact on the distribution of arctic plant spe-
cies (Niittynen et al., 2020; Rissanen et al., 2021). However, the impact 
of the above variables on red-listed plant species need to be assessed in 
the future studies for better clarification. 
The correlations between species occurrence points and environ-
mental variable values are not always apparent (e.g. Araújo et al., 2019). 
SDM tends to overestimate the number of suitable habitats (i.e. false- 
positive occurrences error) (e.g. Svenning et al., 2008; Syfert et al., 
2014; Araújo et al., 2019). To overcome this potential error, historical 
phylogeographic factors, anthropogenic effects (e.g. deforestation), 
ecological competitions, historical disturbance, geographical pattern, 
and non-equilibrium species dynamics should also be included in the 
models (Guisan and Thuiller, 2005; Svenning et al., 2008; Syfert et al., 
2014). However, it is worth to note that few of the variables were not 
available in our analysis such as historical disturbance, ecological 
competition. 
4.3. Implications for biodiversity conservation 
Our model results can be utilized to locate potential new areas for 
red-listed species and to reassess their red-listed status (Syfert et al., 
2014; McKerrow et al., 2018), which serves the data-driven evaluation 
of conservation status endorsed by the IUCN, for example. The models 
also provide information that helps to interpret the ecology of red-listed 
plant species (Di Marco et al., 2017). The maps provide objective evi-
dence for biodiversity conservation without the probably subjective bias 
in maps where the habitat suitabilities are developed using expert 
opinion (Fourcade et al., 2013; Syfert et al., 2014). 
The hotspot maps were produced by summing the habitat suitability 
maps of individual species, which may overestimate hotspot values, 
because the maps do not take into account constraints such as compe-
tition and the dispersal limits of species (Bellamy et al., 2013; Tehrani 
et al., 2020). Furthermore, red-listed plant species usually have small 
populations and may be affected by local extinctions. They may there-
fore be missing in some areas, despite the suitable environmental vari-
ables they contain (Syfert et al., 2014). 
The hotspot maps of red-listed plant species offer a potential tool for 
measuring the overall red-listed plant diversity for the entire country. 
The approach has been generally proven efficient in conservation 
biology (McKerrow et al., 2018; Orme et al., 2005; Tehrani et al., 2020). 
Our models can also be used to extract cold spot maps. These are the 
reverse of hotspots, in that they reveal areas unsuitable for red-listed 
species due to the undesirable condition of crucial characteristics such 
as drainage or large human settlements. Cold spot maps generally show 
a large number of habitats, and like hotspots, cold spots need to be 
monitored to evaluate whether conservation or restoration efforts would 
improve their suitability for red-listed species. 
One way to use the prediction models is to include the historical land 
use and management impacts to help evaluate whether the area of 
hotspots in the southern region would be larger than the present if the 
land use were not as heavy as it is now. A related approach was used in 
Tolvanen et al. (2020), in which the drainage level was theoretically 
altered between 0% and the present level to observe the impact on the 
distribution of red-listed species. Nevertheless, habitat variables other 
than drainage remained unaltered in their assessment. Hence, a question 
remains concerning the total impact of land use on the historic distri-
bution and hotspots of red-listed species until the present. Finding an-
swers to this difficult question would help evaluate the real need for the 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
13
conservation and restoration of habitats for red-listed species. 
5. Conclusions 
This study demonstrated the SDM that incorporates national and 
regional modeling approaches. Three significant conclusions are rele-
vant. First, the distribution of environmental variable values differed 
between the three regions. Second, both national and regional models 
had similar performances, which indicates that realistic species distri-
bution maps can be generated. Last, the transferability (an external 
measure of accuracy) of the regional models (northern, middle, and 
southern regions) was relatively low, which indicates that a good 
regional model does not automatically ensure equally good performance 
when they are transferred to a new case study area. Our study corrob-
orates that transferring SDMs for red-listed plants into unsurveyed areas 
is difficult compared to the internal measure of estimation. To improve 
the transferability of the regional models, we suggest the normalizing of 
environmental variable values. Although we used the maxent algorithm 
for statistical model building, other algorithms might be interesting, 
including deep learning algorithms. Our future research work will lie in 
this direction. 
Funding 
This research work was funded by the Hydrology LIFE (LIFE16NAT/ 
FI/00058), and the project is coordinated by Metsahallitus, Parks & 
Wildlife Finland, and other beneficiaries. 
CRediT authorship contribution statement 
Parvez Rana: Conceptualization, Data curation, Formal analysis, 
Investigation, Methodology, Resources, Software, Validation, Visuali-
zation, Writing - original draft, Writing - review & editing. Anne Tol-
vanen: Conceptualization, Funding acquisition, Methodology, Project 
administration, Supervision, Validation, Writing - review & editing. 
Declaration of Competing Interest 
The authors declare that they have no known competing financial 
interests or personal relationships that could have appeared to influence 
the work reported in this paper. 
Acknowledgments 
We acknowledged the contribution of volunteers for their invaluable 
support during the data collection of the red-listed species database. We 
are also grateful to the anonymous reviewer for their constructive sug-
gestions and feedback, and to Natural Resources Institute Finland for 
their open-access funding. 
Appendix A. Supplementary data 
Supplementary data to this article can be found online at https://doi. 
org/10.1016/j.ecolind.2021.107950. 
References 
Aapala, K., Heikkila, R., Lindholm, T., 1998. Protecting the diversity of Finnish mires. In: 
Vasander, H. (Ed.), Mires of Finland. Finnish Peatland Society, Helsinki, pp. 45–57. 
Ahti, T., Hamet-Ahti, L., Jalas, J., 1968. Vegetation zones and their sections in 
northwestern Europe. Ann. Bot. Fenn. 5, 169–211. 
Allouche, O., Tsoar, A., Kadmon, R., 2006. Assessing the accuracy of species distribution 
models: prevalence, kappa and the true skill statistic (TSS). J. Appl. Ecol. 43, 
1223–1232. https://doi.org/10.1111/j.1365-2664.2006.01214.x. 
Alvarez-martinez, J.M., Suarez-Seoane, S., Palac, C., Sanz, J., Alonso, J.C., 2015. Can 
Eltonian processes explain species distributions at large scale? A case study. Divers. 
Distrib. 21, 123–138. https://doi.org/10.1111/ddi.12256. 
Andrew, M.E., Fox, E., 2020. Modelling species distributions in dynamic landscapes : the 
importance of the temporal dimension. J. Biogeogr. 47, 1510–1529. https://doi.org/ 
10.1111/jbi.13832. 
Araújo, M.B., Pearson, R.G., Thuiller, W., Erhard, M., 2005. Validation of species–climate 
impact models under climate change. Global Change Biol. 11(9), 1504–1513. 
https://doi. org/10.1111/j.1365-2486.2005.01000.x. 
Araújo, M.B., Anderson, R.P., Barbosa, A.M., Beale, C.M., Dormann, C.F., Early, R., 
Garcia, R.A., Guisan, A., Maiorano, L., Naimi, B., O’Hara, R.B., Zimmermann, N.E., 
Rahbek, C., 2019. Standards for distribution models in biodiversity assessments. Sci. 
Adv. 5, 1–12. https://doi.org/10.1126/sciadv.aat4858. 
Arenas-Castro, S., Gonçalves, J.F., Moreno, M., Villar, R., 2020. Projected climate 
changes are expected to decrease the suitability and production of olive varieties in 
southern Spain. Sci. Total Environ. 709, 136161. https://doi.org/10.1016/j. 
scitotenv.2019.136161. 
Barbet-Massin, M., Jiguet, F., Albert, C.H., Thuiller, W., 2012. Selecting pseudo-absences 
for species distribution models: how, where and how many? Methods Ecol. Evol. 3, 
327–338. https://doi.org/10.1111/j.2041-210X.2011.00172.x. 
Bellamy, C., Scott, C., Altringham, J., 2013. Multiscale, presence-only habitat suitability 
models: fine-resolution maps for eight bat species. J. Appl. Ecol. 50 (4), 892–901. 
https://doi.org/10.1111/1365-2664.12117. 
Benito, B.M., Svenning, J., Kellberg-nielsen, T., Riede, F., Gil-romera, G., Mailund, T., 
Kjaergaard, P.C., Sandel, B.S., 2017. The ecological niche and distribution of 
Neanderthals during the Last Interglacial. J. Biogeogr. 51–61 https://doi.org/ 
10.1111/jbi.12845. 
Beven, K.J., Kirkby, M.J., 1979. A physically based, variable contributing area model of 
basin hydrology. Hydrol. Sci. Bull. 24, 43–69. https://doi.org/10.1080/ 
02626667909491834. 
Botella, C., Joly, A., Bonnet, P., Monestiez, P., Botella, C., Joly, A., Bonnet, P., 
Monestiez, P., Munoz, F., Distribution, S., et al., 2018. A deep learning approach to 
species distribution modelling. Multimed. Tools Appl. Environ. 169–199. 
Boyce, M.S., Vernier, P.R., Nielsen, S.E., Schmiegelow, F.K.A., 2002. Evaluating resource 
selection functions. Ecol. Modell. 157, 281–300. 
Brown, J.L., Bennett, J.R., French, C.M., 2017. SDMtoolbox 2.0: The next generation 
Python-based GIS toolkit for landscape genetic, biogeographic and species 
distribution model analyses. PeerJ 5, e4095. 
Byg, A., Martin-Ortega, J., Glenk, K., Novo, P., 2017. Conservation in the face of 
ambivalent public perceptions – The case of peatlands as ’the good, the bad and the 
ugly’. Biol. Conserv. 206, 181–189. https://doi.org/10.1016/j.biocon.2016.12.022. 
CBD, 2020. Zero Draft of post-2020 biodiversity framework, Secretariat of the 
Convention on Biological Diversity, CBD/WG2020/2/3. Kunming, China.  
Cianfrani, C., Lay, L., Hirzel, A.H., Loy, A., 2010. Do habitat suitability models reliably 
predict the recovery areas of threatened species? J. Appl. Ecol. 47, 421–430. 
Connor, T., Hull, V., Vi~na, A., Shortridge, A., Tang, Y., Zhang, J., Wang, F., 2018. Effects 
of grain size and niche breadth on species distribution modeling. Ecography 41, 
1270–1282. https://doi.org/10.1111/ecog.03416. 
Cris, R., Buckmaster, S., Bain, C., Reed, M. (Eds.), 2014. Global Peatland Restoration 
demonstrating SUCCESS. IUCN UK National Committee Peatland Programme, 
Edinburgh.  
Datta, A., Schweiger, O., Kühn, I., 2020. Origin of climatic data can determine the 
transferability of species distribution models. NeoBiota 59, 61–76. https://doi.org/ 
10.3897/neobiota.59.36299. 
Davis, J.C., 1996. Introduction to statistical pattern recognition. Comput. Geosci. 7 (22), 
833–834. 
Deneu, B., Servajean, M., Bonnet, P., Botella, C., Munoz, F., Joly, A., 2021. Convolutional 
neural networks improve species distribution modelling by capturing the spatial 
structure of the environment. PLoS Comput. Biol. 17, 1–21. https://doi.org/ 
10.1371/journal.pcbi.1008856. 
Di Marco, M., Watson, J.E.M., Possingham, H.P., Venter, O., 2017. Limitations and trade- 
offs in the use of species distribution maps for protected area planning. J. Appl. Ecol. 
54, 402–411. https://doi.org/10.1111/1365-2664.12771. 
Duque-Lazo, J., Van Gils, H.A., Groen, T.A., Navarro-Cerrillo, R.M., 2016. Transferability 
of species distribution models: the case of Phytophthora cinnamomi in Southwest 
Spain and Southwest Australia. Ecol. Model. 24, 62–70. 
Elith, J., Leathwick, J.R., 2009. Species distribution models: ecological explanation and 
prediction across space and time. Annu. Rev. Ecol. Evol. Syst. 40, 677–697. https:// 
doi.org/10.1146/annurev.ecolsys.110308.120159. 
Elith, J., Phillips, S.J., Hastie, T., Dudík, M., Chee, Y.E., Yates, C.J., 2011. A statistical 
explanation of MaxEnt for ecologists. Divers. Distrib. 17, 43–57. https://doi.org/ 
10.1111/j.1472-4642.2010.00725.x. 
ESRI, 2018. ArcGIS Desktop: Release 10. Environmental Systems Research Institute, 
Redlands, CA.  
European commision, 2020. EU Biodiversity Strategy for 2030, Bringing nature back into 
our lives. https://eur-lex.europa.eu/legal-content/EN/ALL/?uriˆCELEX: 
52020DC0380. 
Fernandes, R.F., Scherrer, D., Guisan, A., 2019. Effects of simulated observation errors on 
the performance of species distribution models. Divers. Distrib. 25 (3), 400–413. 
https://doi.org/10.1111/ddi.12868. 
Finnish Forest Research Institute, 2014. Finnish statistical yearbook of forestry 2014. 
Vantaa. 
Fourcade, Y., Engler, J.O., Besnard, A.G., Rodder, D., Secondi, J., 2013. Confronting 
expert-based and modelled distributions for species with uncertain conservation 
status : a case study from the corncrake (Crex crex). Biol. Conserv. 167, 161–171. 
https://doi.org/10.1016/j.biocon.2013.08.009. 
Fukunaga, K., 2013. Introduction to statistical pattern recognition. Elsevier. 
Gonzalez and Woods, 2017. Digital Image Processing, 4th edition. Pearson. 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
14
Guisan, A., Thuiller, W., 2005. Predicting species distribution: Offering more than simple 
habitat models. Ecol. Lett. 8, 993–1009. https://doi.org/10.1111/j.1461- 
0248.2005.00792.x. 
Hanski, I., 2015. Habitat fragmentation and species richness. J. Biogeogr. 42, 989–993. 
https://doi.org/10.1111/jbi.12478. 
Harpenslager, S.F., van den Elzen, E., Kox, M.A.R., Smolders, A.J.P., Ettwig, K.F., 
Lamers, L.P.M., 2015. Rewetting former agricultural peatlands: Topsoil removal as a 
prerequisite to avoid strong nutrient and greenhouse gas emissions. Ecol. Eng. 84, 
159–168. https://doi.org/10.1016/j.ecoleng.2015.08.002. 
Heikkinen, R.K., Marmion, M., Luoto, M., 2012. Does the interpolation accuracy of 
species distribution models come at the expense of transferability? Ecography 35, 
276–288. 
Heikkinen, K., Karppinen, A., Karjalainen, S.M., Postila, H., Hadzic, M., Tolkkinen, M., 
Marttila, H., Ihme, R., Kløve, B., 2018. Long-term purification efficiency and factors 
affecting performance in peatland-based treatment wetlands: an analysis of 28 peat 
extraction sites in Finland. Ecol. Eng. 117, 153–164. https://doi.org/10.1016/j. 
ecoleng.2018.04.006. 
Hengl, T., Sierdsema, H., Radovic, A., Dilo, A., 2009. Spatial prediction of species’ 
distributions from occurrence-only records: combining point pattern analysis, ENFA 
and regression-kriging. Ecol. Modell. 220, 3499–3511. https://doi.org/10.1016/j. 
ecolmodel.2009.06.038. 
Higgins, S.I., Larcombe, M.J., Beeton, N.J., Conradi, T., Nottebrock, H., 2020. Predictive 
ability of a process-based versus a correlative species distribution model. Ecol. Evol. 
10 (20), 11043–11054. 
Hirzel, A.H., Le Lay, G., Helfer, V., Randin, C., Guisan, A., 2006. Evaluating the ability of 
habitat suitability models to predict species presences. Ecol. Modell. 199 (2), 
142–152. https://doi.org/10.1016/j.ecolmodel.2006.05.017. 
Hollander, M., Wolfe, D.A., 1973. Nonparametric Statistical Methods. John Wiley & 
Sons, New York.  
Huang, Q., Fleming, C.H., Robb, B., Lothspeich, A., Songer, M., 2018. How different are 
species distribution model predictions?—Application of a new measure of 
dissimilarity and level of significance to giant panda Ailuropoda melanoleuca. Ecol. 
Inf. 1 (46), 114–124. 
Hyvarinen, E., Juslen, A., Kemppainen, E., Uddstrom, A., Liukko, U.M. (Eds.), 2019. The 
2019 Red List of Finnish Species. Ymparistoministerio & Suomen ymparistokeskus, 
Helsinki, p. 704. 
IPBES, 2018. In: The IPBES assessment report on land degradation and restoration. 
Secretariat of the Intergovernmental Science-Policy Platform on Biodiversity and 
Ecosystem Services, Bonn, Germany, p. 744. https://doi.org/10.4324/ 
9781315640051-105. 
Iturbide, M., Bedia, J., Herrera, S., del Hierro, O., Pinto, M., Gutierrez, J.M., 2015. 
A framework for species distribution modelling with improved pseudo-absence 
generation. Ecol. Modell. 312, 166–174. https://doi.org/10.1016/j. 
ecolmodel.2015.05.018. 
Jimenez-Valverde, A., Lobo, J.M., 2007. Threshold criteria for conversion of probability 
of species presence to either-or presence-absence. Acta Oecologica 31, 361–369. 
https://doi.org/10.1016/j.actao.2007.02.001. 
Kappal, S., 2019. Data normalization using median median absolute deviation MMAD 
based Z-score for robust predictions vs. min–max normalization. London J. Res. Sci. 
Nat. Formal. 19, 39–44. 
Kearney, M., Porter, W., 2009. Mechanistic niche modelling: combining physiological 
and spatial data to predict species’ ranges. Ecol. Lett. 12 (4), 334–350. https://doi. 
org/10.1111/j.1461-0248.2008.01277.x. 
Kohonen, T., 2001. Self-Organizing Maps, 3rd edition. Springer-Verlag, Berlin, Germany.  
Kramer-Schadt, S., Niedballa, J., Pilgrim, J.D., Schroder, B., Lindenborn, J., 
Reinfelder, V., Stillfried, M., Heckmann, I., Scharf, A.K., Augeri, D.M., Cheyne, S.M., 
Hearn, A.J., Ross, J., Macdonald, D.W., Mathai, J., Eaton, J., Marshall, A.J., 
Semiadi, G., Rustam, R., Bernard, H., Alfred, R., Samejima, H., Duckworth, J.W., 
Breitenmoser-Wuersten, C., Belant, J.L., Hofer, H., Wilting, A., 2013. The importance 
of correcting for sampling bias in MaxEnt species distribution models. Divers. 
Distrib. 19, 1366–1379. https://doi.org/10.1111/ddi.12096. 
Lentini, P.E., Stirnemann, I.A., Stojanovic, D., Worthy, T.H., Stein, J.A., 2018. Using 
fossil records to inform reintroduction of the kakapo as a refugee species. Biol. 
Conserv. 217, 157–165. https://doi.org/10.1016/j.biocon.2017.10.027. 
Li, J., Fan, G., He, Y., 2020. Predicting the current and future distribution of three Coptis 
herbs in China under climate change conditions, using the MaxEnt model and 
chemical analysis. Sci. Total Environ. 698, 134141. https://doi.org/10.1016/j. 
scitotenv.2019.134141. 
Li, W., Guo, Q., Elkan, C., 2011. Can we model the probability of presence of species 
without absence data? Ecography 34, 1096–1105. https://doi.org/10.1111/j.1600- 
0587.2011.06888.x. 
Lier, M., Korhonen, K.T., Packalen, T., Sauvula-Seppala, T., Tuomainen, T., Viitanen, J., 
Mutanen, A., Vaahtera, E., Hyvarinen, J., 2019. Finland’s forests 2019. Helsinki, 
Natural Resources Institute Finland (Luke).  
Liu, C., Newell, G., White, M., 2016. On the selection of thresholds for predicting species 
occurrence with presence-only data. Ecol. Evol. 6, 337–348. https://doi.org/ 
10.1002/ece3.1878. 
Liu, C., White, M., Newell, G., 2013. Selecting thresholds for the prediction of species 
occurrence with presence-only data. J. Biogeogr. J, 778–789. https://doi.org/ 
10.1111/jbi.12058. 
Low, B.W., Zeng, Y., Tan, H.H., Yeo, D.C., 2021. Predictor complexity and feature 
selection affect Maxent model transferability: evidence from global freshwater 
invasive species. Divers. Distrib. 27, 497–511. 
LUKE, 2019. Finnish statistical yearbook of forestry 2019. Natural Resources Institute 
Finland, Helsinki.  
Manzoor, S.A., Griffiths, G., Lukac, M., 2018. Species distribution model transferability 
and model grain size – finer may not always be better. Sci. Rep. 1–9 https://doi.org/ 
10.1038/s41598-018-25437-1. 
Marcer, A., Saez, L., Molowny-horas, R., Pons, X., Pino, J., 2013. Using species 
distribution modelling to disentangle realised versus potential distributions for rare 
species conservation. Biol. Conserv. 166, 221–230. https://doi.org/10.1016/j. 
biocon.2013.07.001. 
Marsaglia, G., Tsang, W.W., Wang, J., 2003. Evaluating Kolmogorov’s distribution. 
J. Stat. Softw. 8 (18), 1–4. https://doi.org/10.18637/jss.v008.i18. 
Martin, G.D., Magengelele, N.L., Paterson, I.D., Sutton, G.F., 2020. South African Journal 
of Botany Climate modelling suggests a review of the legal status of Brazilian pepper 
Schinus terebinthifolia in South Africa is required. South African J. Bot. 132, 95–102. 
https://doi.org/10.1016/j.sajb.2020.04.019. 
Mateo, R.G., Croat, T.B., Felicísimo, A.M., Mu~noz, J., 2010. Profile or group 
discriminative techniques? Generating reliable species distribution models using 
pseudo-absences and target-group absences from natural history collections. Divers. 
Distrib. 16, 84–94. https://doi.org/10.1111/j.1472-4642.2009.00617.x. 
McCune, J.L., 2016. Species distribution models predict rare species occurrences despite 
significant effects of landscape context. J. Appl. Ecol. 53, 1871–1879. https://doi. 
org/10.1111/1365-2664.12702. 
McKerrow, A.J., Tarr, N.M., Rubino, M.J., Williams, S.G., 2018. Patterns of species 
richness hotspots and estimates of their protection are sensitive to spatial resolution. 
Divers. Distrib. 24, 1464–1477. https://doi.org/10.1111/ddi.12779. 
Merow, C., Smith, M.J., Silander, J.A., 2013. A practical guide to MaxEnt for modeling 
species’ distributions : what it does, and why inputs and settings matter. Ecography 
1058–1069. https://doi.org/10.1111/j.1600-0587.2013.07872.x. 
Mikkonen, N., Leikola, N., Lahtinen, A., Lehtomaki, J., Halme, P., 2018. Forest areas 
important for biodiversity in Finland - Final report of the Zonation analyzes of the 
diversity values of wooded habitats, Reports of the Finnish Environment Institute 9/ 
2018. Finnish Environment Institute. 
Niittynen, P., Heikkinen, R.K., Aalto, J., Guisan, A., Kemppinen, J., Luoto, M., 2020. 
Fine-scale tundra vegetation patterns are strongly related to winter thermal 
conditions. Nat. Clim. Chang. 10, 1143–1148. https://doi.org/10.1038/s41558-020- 
00916-4. 
Norris, D., Ferraz, S.F.D.B., Villani, J.P., Galetti, M., 2011. How to not inflate population 
estimates ? Spatial density distribution of white-lipped peccaries in a continuous 
Atlantic forest. Anim. Conserv. 14, 492–501. https://doi.org/10.1111/j.1469- 
1795.2011.00450.x. 
Oke, T.A., Hager, H.A., 2017. Assessing environmental attributes and effects of climate 
change on Sphagnum peatland distributions in North America using single- and 
multi-species models. PLoS ONE 12, 1–16. https://doi.org/10.1371/journal. 
pone.0175978. 
Orme, C.D.L., Davies, R.G., Burgess, M., Eigenbrod, F., Pickup, N., Olson, V.A., 
Webster, A.J., Ding, T.S., Rasmussen, P.C., Ridgely, R.S., Stattersfield, A.J., 
Bennett, P.M., Blackburn, T.M., Gaston, K.J., Owens, I.P.F., 2005. Global hotspots of 
species richness are not congruent with endemism or threat. Nature 436, 
1016–1019. https://doi.org/10.1038/nature03850. 
Paez, A., Farber, S., Wheeler, D., 2011. A simulation-based study of geographically 
weighted regression as a method for investigating spatially varying relationships. 
Environ. Plann. A 43 (12), 2992–3010. 
Parviainen, M., Luoto, M., Ryttari, T., Heikkinen, R.K., 2008. Modelling the occurrence of 
threatened plant species in taiga landscapes: methodological and ecological 
perspectives. J. Biogeogr. 35, 1888–1905. https://doi.org/10.1111/j.1365- 
2699.2008.01922.x. 
Parviainen, M., Marmion, M., Luoto, M., Thuiller, W., Heikkinen, R.K., 2009. Using 
summed individual species models and state-of-the-art modelling techniques to 
identify threatened plant species hotspots. Biol. Conserv. 142, 2501–2509. https:// 
doi.org/10.1016/j.biocon.2009.05.030. 
Parviainen, M., Zimmermann, N.E., Heikkinen, R.K., Luoto, M., 2013. Using unclassified 
continuous remote sensing data to improve distribution models of red-listed plant 
species. Biodivers. Conserv. 22, 1731–1754. https://doi.org/10.1007/s10531-013- 
0509-1. 
Phillips, S.J., Anderson, R.P., Schapire, R.E., 2006. Maximum entropy modeling of 
species geographic distributions. Ecol. Modell. 190 (3–4), 231–259. 
Phillips, S.J., Dubik, M., 2008. Modeling of species distributions with Maxent: new 
extensions and a comprehensive evaluation. Ecography 31, 161–175. https://doi. 
org/10.1111/j.2007.0906-7590.05203.x. 
Pirinen, P., Simola, H., Juha, A., Kaukoranta, J.-P., Karlsson, P., Ruuhela, R., 2012. 
Climatological statistics of Finland 1981–2010. Finnish Meteorological Institute, 
Reports 2012:1, Finnish Meteorological Institute p83. 
Progin, D., 2018. Modelling habitat suitability of bats in the Prealps. Master thesis. 
Lausanne University, Lausanne, Switzerland.  
Qiao, H., Feng, X., Escobar, L.E., Peterson, A.T., Soberon, J., Zhu, G., Papes¸, M., 2019. An 
evaluation of transferability of ecological niche models. Ecography 42 (3), 521–534. 
R Core Team, 2020. R: A language and environment for statistical computing. R 
Foundation for Statistical Computing, Vienna, Austria http://www.R-project.org/.  
Ranc, N., Santini, L., Rondinini, C., Boitani, L., Poitevin, F., Angerbjorn, A., Maiorano, L., 
2017. Performance tradeoffs in target-group bias correction for species distribution 
models. Ecography 40, 1076–1087. https://doi.org/10.1111/ecog.02414. 
Randin, C.F., Dirnbock, T., Dullinger, S., Zimmermann, N.E., Zappa, M., Guisan, A., 
2006. Are niche-based species distribution models transferable in space? J. Biogeogr. 
33, 1689–1703. https://doi.org/10.1111/j.1365-2699.2006.01466.x. 
Regos, A., Gagne, L., Alcaraz-Segura, D., Honrado, J.P., Domínguez, J., 2019. Effects of 
species traits and environmental predictors on performance and transferability of 
ecological niche models. Scientific reports 9, 4221. 
P. Rana and A. Tolvanen                                                                                                                                                                                                                     
Ecological Indicators 129 (2021) 107950
15
Reisenbüchler, M., Bui, M.D., Rutschmann, P., 2021. Reservoir sediment management 
using artificial neural networks: A case study of the lower section of the alpine 
saalach river. Water 13, 818. https://doi.org/10.3390/w13060818. 
Rissanen, T., Niittynen, P., Soininen, J., Luoto, M., 2021. Snow information is required in 
subcontinental scale predictions of mountain plant distributions. Glob. Ecol. 
Biogeogr. 30, 1502–1513. https://doi.org/10.1111/geb.13315. 
Saarimaa, M., Aapala, K., Tuominen, S., Karhu, J., Parkkari, M., Tolvanen, A., 2019. 
Predicting hotspots for threatened plant species in boreal peatlands. Biodivers. 
Conserv 28 (5), 1173–1204. 
Sallinen, A., Tuominen, S., Kumpula, T., Tahvanainen, T., 2019. Undrained peatland 
areas disturbed by surrounding drainage: a large scale GIS analysis in Finland with a 
special focus on aapa mires. Mires Peat 24, 1–22. 
Senay, S.D., Worner, S.P., Ikeda, T., 2013. Novel three-step pseudo-absence selection 
technique for improved species distribution modelling. PLoS ONE 8, e71218. 
https://doi.org/10.1371/journal.pone.0071218. 
Seppa, H., 2002. Mires of Finland: regional and local controls of vegetation, landforms, 
and long-term dynamics. Fennia-Int. J. Geogr. 180 (1–2), 43–60. 
Shamshirband, S., Jafari Nodoushan, E., Adolf, J.E., Abdul Manaf, A., Mosavi, A., 
Chau, K.W., 2019. Ensemble models with uncertainty analysis for multi-day ahead 
forecasting of chlorophyll a concentration in coastal waters. Eng. Appl. Comput. 
Fluid Mech. 13, 91–101. https://doi.org/10.1080/19942060.2018.1553742. 
Simila, M., Aapala, K., Penttinen, J. (Eds.), 2014. Ecological restoration in drained 
peatlands – best practices from Finland. Metsahallitus, Natural Heritage Services, 
Vantaa 2014, p. 84. 
Skov, F., Svenning, J.C., 2004. Potential impact of climatic change on the distribution of 
forest herbs in Europe. Ecography 27, 366–380. https://doi.org/10.1111/j.0906- 
7590.2004.03823.x. 
Svenning, J., Normand, S., Kageyama, M., 2008. Glacial refugia of temperate trees in 
Europe : insights from species distribution modelling. J. Ecol. 96, 1117–1127. 
https://doi.org/10.1111/j.1365-2745.2008.01422.x. 
Swets, J.A., 1988. Measuring the accuracy of diagnostic systems. Science 240, 
1285–1293. 
Syfert, M.M., Joppa, L., Smith, M.J., Coomes, D.A., Bachman, S.P., Brummitt, N.A., 2014. 
Using species distribution models to inform IUCN Red List assessments. Biol. 
Conserv. 177, 174–184. https://doi.org/10.1016/j.biocon.2014.06.012. 
SYKE, 2009. Finnish Environment Institute spatial drainage data on peatlands. Helsinki. 
Tehrani, N.A., Naimi, B., Jaboyedoff, M., 2020. Toward community predictions: Multi- 
scale modelling of mountain breeding birds’ habitat suitability, landscape 
preferences, and environmental drivers. Ecol. Evol. 10, 5544–5557. 
Thapa, A., Hu, Y., Chandra Aryal, P., Singh, P.B., Shah, K.B., Wei, F., 2020. The 
endangered red panda in Himalayas: potential distribution and ecological habitat 
associates. Glob. Ecol. Conserv. 21, e00890. https://doi.org/10.1016/j.gecco.2019. 
e00890. 
Tikkanen, O.P., Martikainen, P., Hyvarinen, E., Junninen, K., Kouki, J., 2006. Red-listed 
boreal forest species of Finland: associations with forest structure, tree species, and 
decaying wood. Ann. Zool. Fennici 43, 373–383. 
Tolvanen, A., Saarimaa, M., Tuominen, S., Aapala, K., 2020. Is 15% restoration sufficient 
to safeguard the habitats of boreal red-listed mire plant species? Glob. Ecol. Conserv. 
23, e01160. https://doi.org/10.1016/j.gecco.2020.e01160. 
Torres, L.G., Sutton, P.J., Thompson, D.R., Delord, K., Weimerskirch, H., Sagar, P.M., 
Sommer, E., Dilley, B.J., Ryan, P.G., Phillips, R.A., 2015. Poor transferability of 
species distribution models for a pelagic predator, the grey petrel, indicates 
contrasting habitat preferences across ocean basins. PLoS ONE 10 (3), e0120014. 
Trotta-Moreu, N., Lobo, J.M., 2010. Deriving the species richness distribution of 
geotrupinae (Coleoptera: Scarabaeoidea) in Mexico From the overlap of individual 
model predictions. Environ. Entomol. 39, 42–49. https://doi.org/10.1603/ 
EN08179. 
Turunen, J., 2008. Development of Finnish peatland area and carbon storage 1950–2000. 
Boreal Environ. Res. 13, 319–334. 
Valikangas, T., Suomi, T., Elo, L.L., 2018. A systematic evaluation of normalization 
methods in quantitative label-free proteomics. Brief. Bioinformat. 19, 1–11. 
Virzi, T., Lockwood, J.L., Lathrop, R.G., Grodsky, S.M., Drake, D., 2017. Predicting 
American Oystercatcher (Haematopus palliatus) Breeding Distribution in an 
Urbanized Coastal Ecosystem Using Maximum Entropy Modeling. Waterbirds 40, 
104–122. https://doi.org/10.1675/063.040.sp112. 
Wenger, S.J., Olden, J.D., 2012. Assessing transferability of ecological models: an 
underappreciated aspect of statistical validation. Methods Ecol. Evol. 3, 260–267. 
https://doi.org/10.1111/j.2041-210X.2011.00170.x. 
Williams, P., Gibbons, D., Margules, C., Rebelo, A., Humphries, C., Pressey, R., 1996. 
A comparison of richness hotspots, rarity hotspots, and complementary areas for 
conserving diversity of British birds. Conserv. Biol. 10, 155–174. https://doi.org/ 
10.1046/j.1523-1739.1996.10010155.x. 
Wisz, M.S., Guisan, A., 2009. Do pseudo-absence selection strategies influence species 
distribution models and their predictions? An information-theoretic approach based 
on simulated data. BMC Ecol. 9, 1–13. https://doi.org/10.1186/1472-6785-9-8. 
Yu, Z.C., 2012. Northern peatland carbon stocks and dynamics: a review. Biogeosciences 
9, 4071–4085. https://doi.org/10.5194/bg-9-4071-2012. 
Zedler, J.B., Kercher, S., 2005. Wetland resources: status, trends, ecosystem services, and 
restorability. Annu. Rev. Environ. Resour. 30, 39–74. https://doi.org/10.1146/ 
annurev.energy.30.050504.144248. 
Zhang, K., Yao, L., Meng, J., Tao, J., 2018. Maxent modeling for predicting the potential 
geographical distribution of two peony species under climate change. Sci. Total 
Environ. 634, 1326–1334. https://doi.org/10.1016/j.scitotenv.2018.04.112. 
Zhou, Y., 2020. Real-time probabilistic forecasting of river water quality under data 
missing situation: deep learning plus post-processing techniques. J. Hydrol. 589, 
125164. https://doi.org/10.1016/j.jhydrol.2020.125164. 
Zurell, D., Franklin, J., Konig, C., Bouchet, P.J., Dormann, C.F., Elith, J., Fandos, G., 
Feng, X., Guillera-Arroita, G., Guisan, A., et al., 2020. A standard protocol for 
reporting species distribution models. Ecography 43, 1261–1277. https://doi.org/ 
10.1111/ecog.04960. 
P. Rana and A. Tolvanen