Predicting the potential habitat of Russian- Olive (Elaeagnus angustifolia) in urban landscapes

Russian-olive (Elaeagnus angustifolia) is a species native to southern Europe and central and eastern Asia. This species plays an important role in urban landscape design because of its rapid growth, resistance in harsh climates and tolerance to human-caused pressure. Understanding its potential dispersal and restricting parameters are the first steps toward the sustainable use of this species. Here, we used Species Distribution Models to predict the potential distribution of Russian-olive in Iran climate and estimate the possible limiting factors for its spread. Our results highlighted the importance of environmental variables including climatic factors, soil, and lithology in the distribution of this species throughout the country. According to these results, suitable habitats for Russian-olive are located in the north of Iran along the Alborz and Koppeh-Dagh mountain ranges. Therefore, the suitable habitats for this species are limited to only nine percent of the country. A habitat suitability map can be used to evaluate future developments in urban areas and predict the dispersal range of Russian-olive in Iran. Our results show that Russian-olive can be used to create new green spaces in urban climates in the northern regions of Iran.


INTRODUCTION
The Middle East and North Africa are home to five percent of Earth's human population. However, only one percent of the global freshwater resources is located in Middle Eastern and North African countries (Djuma et al. 2016). As a result, water scarcity looms large across the region (Al-Ansari and Knutsson 2011; Al-Ansari et al. 2014;Abbas et al. 2018). To complicate the problem even further, population growth and political tensions threaten the sustainability of existing water resources in the Middle East and North Africa (Djuma et al. 2016).
Consequently, making use of different water sources and enhancing the resilience of water supply is crucial to meet the needs of the increasing urban population (Bichai et al. 2015). The environmental damage associated with urban devel-opment has drawn attention to the need for green spaces in cities, which will lead to increased water use (Zhang et al. 2017). Green spaces are among the indicators of sustainable urban development. When planning for urban green spaces, numerous elements, such as economic, political, social, and cultural factors, along with management and planning considerations need to be taken into account (Haq 2011). Conservation of biological resources and maintaining soil and water quality are among the services provided by urban green spaces (Haq 2011(Haq , 2015. Many studies indicate that plant particularly trees can improve the urban microclimate and influence thermal comfort in various ways including shading, controlling the humidity, wind break, pollutant absorption and produce oxygen (Abreu-Harbich et al. 2015;Thoma et al. 2016;Afshar et al., 2018).
In arid regions such as the Middle East, design of urban green spaces is one of the main challenges facing city planners and urban architects. One solution to address this challenge is the use of native plant species which are adapted to the dry conditions of the region (Katz and Shafroth 2003;Kiseleva and Chindyaeva 2011).
The first step in utilizing native species is identification of their habitat requirements. Species distribution models (SDMs) trace their origin to the 1970s and have remained a common tool for ecologists throughout the following decades (e.g., Guisan and Zimmermann 2000;Guisan and Thuiller 2005;Rooper et al. 2016). In the time since their conception, several SDM algorithms have been developed, as discussed by Elith andLeathwick (2009) andFarashi andAlizadeh-Noughani (2018). These algorithms distinguish the major variables that determine a species' suitable habitat and show how predictor variables impact response variables. Furthermore, SDM algorithms enable researchers to see species' potential distribution (Liang and Stohlgren, 2011;Liang et al. 2017). Through modifications, these algorithms have been optimized for use in fields such as biogeography, ecology, evolution, and species conservation and management (Mikolajczak et al., 2015;Hannah et al., 2015). SDMs have also been used to assess the potential distribution of plant species (e.g., Kumar and Stohlgren 2009;Hemsing and Bryn 2012;Zhang et al., 2013;Guida et al. 2014;Hu et al. 2018). In the present study, we have used SDMs to predict the spatial distribution of Russian-olive (Elaeagnus angustifolia), a native plant species in Iran. Iran is a Middle Eastern country located on Earth's arid belt with upwards 60% of the country's area having an arid or semi-arid climate. In areas that receive little precipitation and experience severe fluctuations from year to year, agriculture is often limited by water availability (Modarres and da Silva 2007).
Russian-olive is native to Eurasia that occurs on coasts, in riparian areas, along watercourses, in other rela-tively moist habitats and also in many arid and semiarid regions of the world (Klich, 2000;Peterson et al., 2003). Soil salinity (low to medium concentrations), pH and water supply and moisture (low) are important environmental factors in Russian-olive habitat (Carman, 1982;Zitzer and Dawson, 1992;Reynolds and Cooper, 2010;Dubovyk et al., 2016). Russian-olive is resistant to drought (+46 °C) and frost (-46 °C) (Stratu et al., 2016;Akbolat et al., 2008). This tree is an ecologically valuable plant that are adapted to a variety of harsh conditions such as cold, drought, and salinity or alkalinity of soil (Asadiar et al. 2013;Zhang et al. 2018). The species endures through water scarcity by using groundwater (Katz and Shafroth 2003). Along with its desirable ecological characteristics, Russian-olive possess aesthetic values such as its beautiful oval crown, arching branches, silver leaves and shiny dark red fruits. Therefore E. angustifolia is particularly suitable for urban landscapes in arid regions such as Iran. This tree can be used to create sustainable green spaces in urban climates of Iran.

Study area and species
Iran is located in Western Asia between 24˚-40˚ N and 44˚-64˚ E. Due to its habitat diversity and phytogeographic variety, Iran hosts rich biodiversity. Over 8,000 species of plants are found in Iran, of which 1,810 are endemic (Ghahraman and Attar 2000; Willis 2001). Russian-olive is a deciduous tree, sometimes with a shrubby habit, in the family Elaeagnaceae (Saboonchian et al. 2014). This species naturally grows in central and eastern Asia and southern Europe. Russian-olive grows quickly, reaching a maximum height of 10 m and maximum trunk diameter of 30 cm. Trees usually bear fruit after 5-6 years (Katz and Shafroth 2003).

Species distribution models
SDMs were developed in Biomod2 package (Thuiller et al. 2009(Thuiller et al. , 2014 in R version 3.1.25 (R Core Team 2014). 10 different algorithms were used to study the species (Tab. 1). The algorithms can be categorized as: regression, machine learning, classification and enveloping algorithms. Regression-based algorithms include generalized linear models (GLMs) and generalized additive models (GAMs) which generate linear and non-linear equations between presence data and environmental variables, respectively. Machine learning algorithms include artificial neural networks (ANN), boosted regression trees, (BRT), multivariate adaptive regression splines (MARS), maximum entropy (MaxEnt), and random forest (RF). Machine learning algorithms directly generate the environmental space using input data. Classification algorithms such as classification and regression trees (CART) and flexible discriminate analyses (FDA) successively divide data into homogenous partitions. Surface range envelope (SRE), the only enveloping method used in this study, investigates environmental conditions at the points of occurrence and uses the results to find similar areas (Merow et al. 2014).
Variable importance was calculated by a permutation procedure used in biomod, which is independent of the modelling technique. Once the models were trained (i.e., calibrated), a standard prediction was made. Then, one of the variables was randomized and a new prediction was made. The correlation score between the new prediction and the standard prediction was calculated and gave an estimation of the variable importance in the models (Thuiller et al., 2009).
Models were evaluated using the True Skill Statistic (TSS). TSS is the sum of sensitivity and specificity minus 1, and does not depend on prevalence (Allouche et al. 2006;Fielding and Bell 1997). TSS was used to create an ensemble-forecasting framework, as per Araújo and New (2007). All models contributed to the ensemble model. However, those with better performance, as indicated by TSS, were given more weight (Thuiller et al. 2009). A threshold value was defined by maximizing training sensitivity and specificity in order to create a binary (presence/absence) map from outputs of the algorithms (Liu et al. 2005;Liu et al. 2011). Sensitivity and specificity are statistical index of the performance of a binary classification analysis. Sensitivity calculate the proportion of actual presences which are correctly predicted as such, while specificity calculate the proportion of pseudoabsences which are predicted as absences. By maximizing the sum of sensitivity and specific-ity, the associated threshold corresponds to the point on the ROC curve (i.e. sensitivity against 1-specificity) whose tangent slope is equal to 1 (Kaivanto 2008;Jiguet et al. 2011). The approach was selected to calculate the threshold for presence/absence predictions in biomod2 (Liu et al. 2005).

Presence data and environmental variables
Occurrence records and distribution of the species were obtained from herbariums of Ferdowsi University of Mashhad, Tehran University, and University of Birjand. Flora Iranica (Rechinger, 1963(Rechinger, -2015 and Flora of Iran (Assadi et al. 1988(Assadi et al. -2017. Herbaria data were obtained from field samplings between 2009 and 2019. The coordinates of all the occurrence points were recorded using a hand-held multichannel Global Positioning System (GPS) receiver with a positional accuracy of ±5 m. The spatially correlated presence points were removed using spatial autocorrelation and Moran's I test. The number of presence points was 83 (Fig. 1).
Topographic, geographic, edaphic, and climatic variables were used as input for the algorithms. Topographic variables were obtained from the national cartographic center of Iran (NCC) at 1-km spatial resolution. Geological survey and mineral exploration of Iran (GSI) provided the geographic data at 1-km spatial resolution. Edaphic variables were accessed from the agricultural research, education and extension organization of Iran (AREEO) at 1-km spatial resolution.
Mean elevation and mean slope for all raster cells in a 1-km radius were the two topographic variables used in modeling. Geographic and edaphic variables included soil orders and lithology, respectively. An initial set of 20 climatic variables, including precipitation, temperature, and solar radiation were obtained from the Worldclim database (http://www.worldclim.org). Climatic variables were used at a resolution of 30'' (~ 1km). The correlation between all pairs of variables was tested. If -0.7 > r > +0.7, one of the two variables was excluded from the input data. The correlation tests reduced the number of variables to 12, which were subsequently used to model habitat suitability (Tab. 2).

RESULTS
All ten models showed a relatively good performance predicting the distribution of Russian-olive (Tab. 1). The results of modeling evaluation based on the TSS values showed that the combination of models performed relatively better than each individual model. Moreover, a model evaluation test showed that ensemble model performed better than other distribution models. The distribution map obtained from the ensemble model has been presented in Fig. 1. Our results showed that most of the suitable habitats for Russian-olive are located in the north of Iran. Only 9.5 percent of the country was suitable to grow this species (Fig. 1).
Suitable habitats based for each province have been presented in a separate map (Fig. 2). North Khorasan had the highest, and Ilam and Bushehr had the lowest proportion of suitable habitats among all provinces (Fig. 2). The relative importance of environmental variables changed based on different models. According to ensemble model, the most important environmental variables to predict habitat suitability for this species were lithology (50% of the contribution), mean temperature of the warmest quarter (22% of the contribution), annual solar radiation (10% of the contribution) and soil order (8% of the contribution) (Tab. 2). Response curves for the four dominant environmental factors are shown in Fig. 3. There are unimodal relationships between habitat suitability and annual solar radiation. Peak presence probability was observed at 8150 kJ m -2 day -1 . The relationship between the habitat suitability values and mean temperature of the warmest quarter was best described by an exponential decay with the peak response at 5-7 °C. The results also demonstrated that any increase in mean temperature of the warmest quarter and annual solar radiation led to a decrease in habitat suitability for Russian-olive. The relationship between the habitat suitability values with soil order and lithology showed that this species could grow in different soil and rock classes. However, the highest presence probability is observed in rocky lands and high-level piedmont fan and valley terrace deposits (Fig. 3). DISCUSSION Iran is a large country, containing a variety of climates. While the northern regions have a temperate climate, southern regions are dry and frequently experience droughts and water scarcity (Abbaspour et al., 2009;Bannayan et al., 2010). Our results show the prominent role of mean temperature of warmest quarter, annual solar radiation, lithology, and soil order in creating a suitable habitat for Russian-olive. The contribution of other variables was not considerable. Previous studies have shown that Russian-olive is capable of growing under both flooded and drought conditions in its native range (Asadiar, et al., 2013, Stannard et al., 2002 as well as its introduced range (Katz and Shafroth, 2003;Reynolds and Cooper, 2010). E. angustifolia's extensive root network allows it to utilize moisture stored in deep soil or groundwater (Cui et al., 2015;Dubovyk et al., 2016). Owing to insufficient hydro-geological data, we could not use these variables in our study. Nevertheless, we recommend including them in future studies when they become available for Iran.
Our findings also reveal the importance of environmental variables such as soil (soil orders) and lithology in determining suitable habitats for Russian-olive, which supports the findings of previous studies (Zitzer and Dawson, 1992;Carman and Brotherson, 1982;Khamzina et al., 2009;Collette and Pither, 2015). The results demonstrate how Russian-olive can survive only under certain climatic conditions but can continue to grow on a number of soil orders and lithological formations (Lesica and Miles 2001;Katz and Shafroth, 2003;Reynolds and Cooper 2010;Collette and Pither, 2015). This makes Russian-olive a good candidate for shelterbelts in different regions (Olson and Knopf 1986;Pearce et al., 2009).
Roughly 9% of Iran is suitable habitat for Russian-olive, stretching along the Alborz and Koppeh-Dagh mountain ranges (Fig. 1). The Alborz and Koppeh-Dagh are comparable with temperate European mountain ranges such as the Alps in terms of endemism (Tribsch and Schonswetter 2003;Noroozi et al. 2008Noroozi et al. , 2018. Iranian provinces vary regarding habitat suitability for Russian-olive. All provinces, with the exception of Ilam and Bushehr (in the west and south of Iran, respectively), contained suitable habi-tats for Russian-olive. North Khorasan (64.7%), Qazvin (44.8%), and Alborz (42.4%) had the highest proportion of suitable habitats for Russian-olive. Suitability maps can inform future urban development and predict the future range of Russian-olive.
Therefore, it is suggested to protect the critical habitats of Russian-olive and use this species in urban green spaces. Russian-olive is not a demanding species and can survive for 50-80 years in different conditions. E. angustifolia is used to as a soil stabilizer, a hedge plant, and a fragrant ornamental. Due to its characteristics, Russian-olive is used in shelterbelts and urban landscapes (Kolesnikov, 1974;Kiseleva and Chindyaeva, 2011).
Russian-olive can become invasive (Reynolds and Cooper, 2010;Collette and Pither, 2015). After its introduction as an ornamental plant, Russian-olive became invasive in the US and Canada in the early 20 th century (Katz and Shafroth 2003). The species negatively affected riparian forests and, as a result, was declared a noxious species in Colorado and New Mexico (Katz and Shafroth 2003;Collette and Pither, 2015). Introduction of this species to areas outside its native range should be done with caution. However, such considerations are not needed when planting Russian-olives in its native range since the species will not disrupt the natural processes of its native ecosystems (Strauss et al., 2006;Marsh-Matthews et al., 2011;Zhang et al., 2018). Moreover, native species can be advantageous to the local economy. As a result, we recommend the use of Russian-olive in urban landscapes in northern Iran.
A common assumption among SDMs is that species can only establish in areas that are ecologically similar to their native range (Kearney 2006). However, a species niche might change (Broennimann et al., 2007). As a result, the output of SDM algorithms is an approximation of species' niche in new environments. The differences in bioclimatic conditions between native areas and those we are making predictions for might lead to an underestimation of actual suitable areas. Thus, more accurate predictions can only be made by taking into account both biotic and abiotic variables and their interactions. These studies can be further improved through comparisons with areas under invasion by alien invasive species. In the meantime, the mere presence of suitable habitats for a species should not encourage managers to use the species before more extensive investigations are performed. However, the efficiency of SDMs is affected by several parameters (Allouche et al. 2008) such as the characteristics of environmental data (e.g. type, variance data; Aguirre-Gutiérrez et al. 2013), characteristics of species data (e.g. geographical accuracy, sample size, field survey constraints, or auto-correlation structure; Huettmann and Diamond 2006), species ecology (e.g. distribution range, abundance, niche limits of species; Saupe et al., 2012), computer power (i.e. too many cells may be too demanding on computer resources), model (e.g. presence only/presence-absence; Aguirre and Gutiérrez et al., 2013), and spatial resolution (Farashi and Naderi 2017). Despite their shortcomings, SDMs can still help us grasp the biological history of a species distribution (Silva Rocha et al., 2015). Further investigation is needed to study niche shift, distinguish the most influential variables, and pinpoint the role of other factors in determining distribution of the species.  Pale-red, polygenic conglomerate and sandstone 45 E1f Silty shale, sandstone, marl, sandy limestone, limestone and conglomerate 46 E1l Nummulitic limestone 47 E1m Marl, gypsiferous marl and limestone 48 E1s Sandstone, conglomerate, marl and sandy limestone 49 E2-3f Sandstone, calcareous sandstone and limestone 50 E2c Conglomerate and sandstone 51 E2f Sandstone, calcareous sandstone and limestone 52 E2l Nummulitic limestone 53 E2m Pale red marl, gypsiferous marl and limestone 54 E2mg Gypsiferous marl 55 E2s Sandstone, marl and limestone 56 E2sht Tuffaceous shale and tuff 57 E3c Conglomerate and sandstone 58 E3f Sandstone-shale sequence with siltstone, mudstone, limestone and conglomerate 59 E3m Marl, sandstone and limestone 60 E3sm Sandstone and marl 61 Ea.