Mureş Floodplain Natural Park is located in the western part of Romania (20°53'E; 46°07'N) in the Panonic biogeographic region (Fig. 1). The study-area covers 17,455 ha and overlaps the lower part of the Mureş River (tributary of Tisa River), occupying the embanked enclosure of the river between the city of Arad and the state border with Hungary (Bălteanu et al. 2016).

Figure 1.

Location of the Mureş Floodplain Natural Park in Romania.

The Mureş Floodplain Natural Park, established in 2004 through Government Decision no. 2151/2004, falls into V IUCN category - Protected Landscape (Natural Park) and Natura 2000 European Network, both SPA – Special Protection Areas and SCI - Site of Community Importance. Furthermore, since 2006, the area was included on the RAMSAR list as Wetlands of International Importance. The Park is a typical wetland area with running and still waters, alluvial forests, as well as an important place for nesting and passage for a large number of bird species of international importance (Dumitraşcu et al. 2013), hosting specific wetland habitats and species of conservation interest within four nature reserves: Prundu Mare-Pecica, Igriş Isles, Insula Mare Cenad and Cenad Forest (Bălteanu et al. 2016).

Generally, the area is a floodplain with altitudes decreasing from about 110 m (in east) to about 80 m (in west). The climate is temperate-continental with oceanic humid influences (Bogdan 2004), with almost 10–11 °C mean annual air temperature and 500–600 mm mean annual precipitation (Bogdan et al. 2016). Forests (in the eastern half) and agricultural land (mainly arable) represent the main land use/cover category of the Park.

According to the geographical distribution of habitats, the importance of the species and management, the Park is divided into three different zones: (1) the totally protected area (6%) which includes the most valuable natural elements, (2) the sustainable management area (92%), also called the buffer zone, which makes the transition between the totally protected area and (3) the sustainable development area (2%) which includes built-up areas or natural resources exploitation sites that existed prior to the designation of the protected area. Within the totally protected area, any form of use of natural resources, construction or investments which do not meet the sustainable management of the protected natural area and/or scientific research activities is forbidden. In the buffer zone, it is prohibited to build new constructions, except for those that strictly serve the protected natural area, the scientific research activities or those meant to ensure national safety or the prevention of natural disasters. Investments or development activities are accepted in the sustainable development area, priority being given to tourism, albeit respecting the sustainable use of natural resources and the prevention of any significant negative effects on biodiversity (Mureş Floodplain Natural Park Administration 2016).

The current study aims to explain the relationships between A. fruticosa occurrence and its explanatory driving forces, on one hand and to model the probability of the potential distribution using spatial analysis and binary logistic regression (BLR), on the other. The methodology, used in the present study, includes three main stages: the extraction of the geospatial datasets, the spatial analysis using GIS and the statistical analysis using SPSS (Statistical Package for the Social Sciences) software package. Synthetically, Figure 2 describes the methodology applied to calibrate, simulate and validate the model in order to assess the potential distribution of A. fruticosa in Mureş Floodplain Natural Park.

Figure 2.

The flowchart showing the methodology used to assess A. fruticosa in Mureş Floodplain Natural Park.

Based on field research findings, as well as on data availability, eight spatial datasets representing the dependent and the independent variables (Fig. 3) were employed to model the probability of A. fruticosa’s potential distribution: A. fruticosa presence/absence (AF), soil type (ST), depth to water (DW), proximity to water (PW), Normalised Difference Vegetation Index (NDVI), proximity to near roads (PR), forest fragmentation (FF) and proximity to near settlements (PS). Due to the relative homogenous topographical characteristics with rather insignificant altitudinal difference (about 30 m), as well as to its longitudinal and latitudinal low extension (00°47' and 00°07', respectively), the explanatory factors related to climate and relief were not considered. Moreover, the scale of the study area, coupled with the particular local environmental features, makes this research a valuable local-scale approach, allowing a better understanding of the environmental issues and an increased potential for being included into the decision-making process. Thus, we consider that such an approach can be replicated to other geographical areas with similar environmental conditions or extrapolated from local to regional scales.

Figure 3.

Raster layers representing the dependent and independent variables.

Dependent variable: A. fruticosa occurrence

The layer of A. fruticosa presence/absence in the Mureş Floodplain Natural Park was derived from the data collected during the field surveys which were carried out between 2012 and 2016. The mapping was realised using topographic maps (1:25000 scale) and, for more accuracy, using orthophoto images (scale 1:5000) and GPS measurements. The species was identified in various habitat types where field relevés with over 20% coverage (according to Braun – Blanquet scale) and frequency (according to Raunkiaer scale) were taken into consideration (Grigorescu et al. 2014). Thus, the total mapped area covers a surface of 363 ha, including polygons with different coverage of A. fruticosa.

With respect to the dataset used to model the probability map, a binary raster with the categories “presence” (coded with 1) and “absence” (coded with 0) was generated (Fig. 3), in order to discriminate the cells with A. fruticosa occurrence from the non-occurrence.

Explanatory factors

After processing different maps (topographic, soil and hydrogeological) and orthophoto images, six thematic layers were extracted and analysed: soil characteristics, piezometric level, surface waters, forests, built-up areas and roads from where soil type, depth to water, forest fragmentation, distance to roads, distance to waters and distance to settlements were derived.

In order to assess the occurrence of A. fruticosa within the dominant vegetation cover, the NDVI(Normalised Difference Vegetation Index) was used. This index was calculated using the Landsat 8 OLI_TIRS (Operational Land Imager_Thermal Infrared Sensor, accessed July 18, 2017) acquired from the United States Geological Survey – USGS (portal available at http://www.usgs.gov/). This index, developed by Rouse et al. in 1974 and widely used for the remote sensing of vegetation, is a measure of surface reflectance and gives a quantitative estimation of vegetation growth and biomass (Hall et al. 1995). The NDVI values were obtained by employing the following formula:

, (Eq.1)

where NIR = reflectance in the near infrared band; RED = reflectance in the red (visible) band.

The resultant NDVI values ranged between –0.245 and 0.714. After the visual interpretation of orthophoto images, the NDVI values were considered as representing: no vegetation areas (aquatic surface, bare soils, built-up areas, recent riverbed deposits) (NDVI <0); agricultural crops (NDVI = 0…0.25); herbaceous vegetation (NDVI = 0.25…0.4); transitional woodland-scrub (0.40…0.55); forest vegetation (NDVI >0.55).

Total edge (TE), calculated as the sum of the lengths (m) of all edge segments in a class or landscape (McGarigal and Marks 1995) was used to quantify the forest fragmentation. In this case, the TE was calculated just for the forested class, in a 25 ha window size, using Patch Analyst (Rempel et al. 2012), a GIS tool developed to analyse spatial landscape patches and model the attributes associated with patches, according to numerous statistical metrics. The resultant values were categorised in: slightly fragmented forests (TE <2000 m); moderate fragmented forests (TE = 2000…5000 m); highly fragmented forests (TE >5000 m). Moreover, based on the Euclidean multiple ring buffers, maps depicting the distance to roads (buffer = 0.1 km), waters (buffer = 0.5 km) and settlements (buffer = 0.5 km) were created. Twenty-one binary rasters were generated to distinguish the ten soil type classes, three piezometric levels, three forest fragmentation classes and five NDVI classes. Finally, twenty-four categorical and continuous independent variables were prepared and homogenised as raster with 30×30 m cell (equivalent to the spatial resolution of the Landsat image used to derive NDVI) for further spatial and statistical analyses (Table 1). For these variables, several assumptions related to location suitability were presumed in order to explore the relationship between the invasive species and its explanatory factors.

In the present study, to assess the relationships of the site characteristics and A. fruticosa presence/absence, BLR was applied. This method is the most commonly used parametric model aimed at determining the empirical relationships between a dependent and several independent variables (McCullagh and Nelder 1989), where the dependent variable is a binary presence (1) or absence event (0) and the independent variables are categorical and/or continuous variables. If binary values 1 and 0 are used to represent A. fruticosa occurrence and no occurrence, respectively, the probability of the presence of the species for any specific grid cell was calculated using the logistic curve as described by the logistic function (Kleinbaum 1994):

, (Eq.2)

then the probability of occurrence can be estimated with the following logistic regression model:

, (Eq.3)

where P(Y =1|X₁, X₂, ..., X_i) is the probability of the dependent variable Y being 1 given (X₁, X₂, ..., X_i), i.e. the probability of a cell of being invaded by invasive species; X_i is an independent variable representing the explanatory factors of A. fruticosa and β_i is the coefficient for variable X_i.

The response of these regression functions is visualised into the raster probability map based on the location suitability, given the probability of the occurrence of A. fruticosa in each resultant raster cell.

In order to reduce the effects of multi-collinearity, before the logistic regression analysis, Pearson correlations between each pair of independent variables were conducted and examined. In case of strong correlations (min. ±0.7), the better predictor variable (in univariate trials) was retained. Furthermore, to verify the explanatory power of the variables included in the sub-model, the Cramer’s V statistics tool was used. Cramer’s V is a statistic that transforms chi-square (for a contingency table larger than two rows by two columns) to a range of 0–1, where unit value indicates complete agreement between the two nominal variables (Liebetrau 1983). The test procedure is based on contingency table analysis which can test the strength of the association between the dependent variable and both continuous and categorical independent variables.

The BLR was performed using the backward stepwise method in SPSS in order to obtain the best-fit combination for predictors. Thus, the variables, which collectively best explain A. fruticosa occurrence, were adopted by the regression model. To indicate the effectiveness of the each sets, a Nagelkerke pseudo R square (Nagelkerke 1991) was determined. Furthermore, ROC/AUC (Relative Operating Characteristic/Area Under Curve) was used to test the “goodness of fit” (Pontius and Schneider 2001). In the standard ROC approach, the predictive probability map is compared with the map of the true binary event in order to assess the spatial coincidence between the event and the probability values (Mas et al. 2013). This graph displays the predictive accuracy of the logistic model, which can be evaluated using the area under the ROC curve (AUC). A completely random model gives a ROC value of 0.5, while a perfect fit results in a ROC value of 1.0. For the best-fit combination for resultant predictors, the maximum likelihood estimator (Hosmer and Lemeshow 1989) was determined. In the BLR, the model is considered to fit if the value of the Hosmer-Lemeshow test shows a value higher than p-value (0.05).

Based on the corresponding coefficients of the best fit predictor set, the relative contribution of the explanatory variables of the A. fruticosa occurrence was assessed and the potential distribution probability map was generated. To categorise the resultant map, five classes were used to classify the probability values: very high, high, medium, low and very low probability. The classification was performed by Natural breaks (Jenks), a method that seeks to reduce the variance within classes and to maximise the variance between classes (Jenks 1967), commonly used in GIS techniques for grouping spatial values that are not evenly distributed.

Usually, in the analysis and modelling of spatial data, real datasets are used to validate the performance using different techniques. A typical procedure is splitting data into two parts (Kanevski and Maignan 2004): training set (used to develop the model) and validation set (used to estimate the ability of the model). The proportions of data included in each dataset are somewhat arbitrary and dependent on the total mapped area available, 70% for calibration and 30% for validation being commonly used (Pearson 2010). According to the available datasets representing A. fruticosa occurrence (mapped) and to build the model with a significant percent for the training set, in the present study a 70%/30% training/validation split was considered. Furthermore, to eliminate sampling biases and associated subjectivism, the random-partition was used to extract the data for validation. Then, based on the cross-classification technique, the analysis between validation dataset and probability map was achieved and the frequency of the A. fruticosa occurrence was identified and quantified for each probability class.

In terms of the analysed explanatory factors, A. fruticosa distribution across the park has a relative spatial heterogeneity. The frequency analysis (Fig. 4) shows that A. fruticosa occurs in various conditions but with differences mainly according to the soil type, vegetation cover and distance to roads. In relation to soil type, A. fruticosa overlaps largely protisols and fluvisols (92% of the total mapped area). Similar to the soil particularities, the roads appear to have the most important role in facilitating the establishment of A. fruticosa. In detail, the frequency of species related to distance to nearest roads, calculated for 0.1 km buffer rings, shows that 68% of the mapped area is identified in the first 0.1 km and 91% in the first 0.2 km. In relation to the depth to water, the most significant areas with A. fruticosa (58%) were mapped in the high floodplain, the sand banks and floodplain terraces where the aquifer level is situated at 2.0–5.0 m depth. In relation to the forest fragmentation, A. fruticosa was mainly found (69%) in the moderate fragmented forests (TE = 2000–5000 m/25 ha). Related to NDVI values, the large occurrence of A. fruticoasa (46%) is mainly related to values ranging between 0.40 and 0.55, thus indicating a preference of the invasive species for transitional woodland-scrub. Furthermore, the notable occurrence (25%), in relation to the highest NDVI values (>0.55) and 17% in relation to the medium positive NDVI values (0.25–0.40), indicate also a preference for forest vegetation and grasslands. In relation to the water’s vicinity, 55% of the total mapped areas with A. fruticosa are situated close to the Mureș River, within the first 0.5 km buffer ring. The frequency distribution of A. fruticosa indicates also an increasing occurrence in relation to the distance to settlements. Thus, 43% of the total mapped areas are distributed in the first 1 km and 73% in the first 2 km buffer rings.

Figure 4.

The distribution of A. fruticosa (mapped) in relation to the analysed explanatory factors.

The Pearson correlation analysis between the independent variables showed that the variables were not highly inter-correlated (max. ±0.30), which suggests the absence of multi-collinearity. The highest values were found between NDVI (>0.55) and slightly fragmented forests (0.298), depth to water (<2 m) and protisols (0.255). The lowest coefficients (0.001) were found between slightly fragmented forests and chernozems soil type, moderate fragmented forests and depth to water (>5 m), NDVI (0–0.25) and lacovishte soil type.

The explanatory power of the independents’ variables was tested based on the Cramer’s V statistics. According to this method, the analysed explanatory factors were not strongly associated with A. fruticosa occurrence. Overall, continuous and few categorical variables were found to have better association with A. fruticosa occurrence with Cramer’s V values between 1.5 and 2.3: proximity to roads (V = 0.224), proximity to settlements (V = 0.221), proximity to waters (V = 0.193), piezometric level <2 m (V = 0.151), moderate forest fragmentation (V = 0.161) and NDVI values between 0 and 0.25 (V = 0.152). Furthermore, except for erodisols (V = 0.143), protisols (V = 0.122), fluvisols (V = 0.113), piezometric level between 2 and 5 m (V = 0.114), highly fragmented forests (V = 0.105) and NDVI values between 0.25 and 0.40 (V = 0.11), remaining variables have values less than 0.1, indicating a weakly association with the A. fruticosa occurrence.

Setting the backward stepwise in the BLR, eight steps for the best predictor sets resulted (Table 2). Variables that were not statistically significant, associated with A. fruticosa occurrence within the 95% confidence interval, were identified and automatically excluded by the model. Thus, the best-fit combination for predictors was found in step eight, which includes seventeen explanatory factors (Table 3).

For set-8, the regression “goodness of fit” measured by the Nagelkerke R² is -0.243 which, according to Clark and Hosking (1986), indicates that the model is a good fit for the data. Therewith, Hensher and Johnson (1981) also stated that pseudo R² value between 0.2 and 0.4 can be considered as an extremely good fit. The predictor set-8 also attained the highest accuracy (AUC=0.798) showing a prediction ability of 79.8% of the model. Furthermore, the Hosmer-Lemeshow significance test resulted in Chi-square = 13.39 and p = 0.063 (>0.05), indicating a good fit of the model.

The relative contribution of the explanatory factors was evaluated using the corresponding coefficients in the BLR (Table 3). Based on the coefficients’ values, all the explanatory variables were ranked. Thus, amongst all variables, erodisols, protisols and luvisols were found as the most significant predictors for A. fruticosa occurrence in the study area. All values of OR are greater than one, indicating a higher probability of A. fruticosa occurrence in those areas comparing to other soil type classes. The probability of A. fruticosa occurrence in areas with erodisols is larger than the probability in areas covered with protisols. The areas with protisols present more suitability for A. fruticosa occurrence than areas with luvisols. This can be seen from the odds ratio values of 18.59, 8.17 and 6.96 in a decreasing order for erodisols, protisols and luvisols, respectively.

The regression model showed a positive relationship between A. fruticosa occurrence and depth to water for piezometric level less than 2 m and between 2 and 5 m and negative relation with respect to a piezometric level higher than 5 m. This means that, with the increase in depth to water, the A. fruticosa occurrence decreases due to less connectivity to groundwater. This can be seen in the odds ratio values (4.89, 2.38 and 0.35) in a decreasing order for the piezometric level. Furthermore, the positive values of β and OR show that A. fruticosa tends to spread in moderate and mainly highly fragmented forests. However, the positive values of β (0.19) and the value of OR greater than one (1.21) for slightly fragmented forests also demonstrate that the species can spread in rather compacted afforested areas.

The model also demonstrates that A. fruticosa is in relation to vegetation cover, the positive β coefficients and OR values for NDVI values >0.25 showing the tendency of species to spread in areas with grasslands, transitional woodland-scrub and afforested areas. The estimated β value (–0.694) and OR (0.50) for the proximity to roads indicates that the probability of A. fruticosa occurrence further away from roads is less expected. Specifically, the probability of species occurrence would decrease 2 times if distance to roads increases by 0.1 km. The model demonstrates that A. fruticosa occurrence is not significantly controlled by the proximity to settlements, however the negative value of β (–0.16) indicates that, with the increase in distance to settlements, the probability of this invasive terrestrial species to occur decreases. Thus, the odds of A. fruticosa occurrence in an area 0.5 km closer to settlements is estimated to be 1.17 as large as that in areas further away from settlements.

The regression results for proximity to waters (β = 0.006; OR = 1.006) revealed that they have no significant influence on A. fruticosa occurrence.

The probability of A. fruticosa occurrence was assessed by plugging the β coefficients of the logistic regression model containing the 17 significant predictors (Table 3) into Eq. (3). Thus, the probability map (Fig. 5) indicated that 24.9% of the grid cells have high and very high suitability for A. fruticosa occurrence, largely in the eastern half of the Park, close to Felnac, Pecica and Semlac localities. Here, the suitability is mainly characterised by the favourable soil type, the presence of numerous agricultural and forestry roads and the large extension of the pastures and transitional woodland-scrub vegetation. On the other hand, the lowest probability values are in the western (near Cenad locality), north-eastern and southern parts of the protected area (near Arad and Secusigiu localities) where the unsuitable soil type classes and piezometric level ranking between 2 and 5 m or arable lands and compacted forests are predominant.

Figure 5.

The mapped areas (A) and the probability of A. fruticosa occurrence, based on the BLR (B).

In order to conduct the spatial validation of the model, the map of A. fruticosa occurrence probability, computed using the logistic regression model, was compared with the actual A. fruticosa occurrence (reference datasets used for validation). Thus, the cross-classification map reveals a relatively good spatial fit between the observed data and the predicted data (Fig. 6). Both very high and high probability classes include 69.9% of total cells, representing the real A. fruticosa location used as the validation dataset. Furthermore, only 10% of the total pixels, representing the real A. fruticosa used as the validation dataset, overlap the very low and low probability classes.

Figure 6.

The frequency of A. fruticosa occurrence (datasets used for validation) in the probability classes.

The field surveys in the Mureș Floodplain Natural Park have shown A. fruticosa occurrence in different ecosystems and habitats, significantly affecting the native vegetation. The species were identified mainly along the forest roads edges and forest glades (especially north to Mureș River), as well as along the edges of arable lands (mainly abandoned, unused) with a tendency to invade them in the west and north-west of Felnac locality. The largest area was identified in the western part of Pecica locality and north-west of Fenlac, along the forest roads and at the contact between forested areas and pastures or arable lands. Therewith, the species has a significant spread south-west of Şeitin locality on arable lands and along Mureş River (Fig. 7). Selected biological indices (coverage, frequency, abundance) computed in six representative sites during the field research, indicate a preference mainly for the riparian habitats (with Salix alba and Populus alba), alluvial forest (with Alnus glutinosa and Fraxinus excelsior), riparian mixed forests (with Quercus robur, Ulmus laevis, U. minor, Fraxinus excelsior, F. angustifolia) and muddy banks (with Chenopodium rubri pp and Bidention vegetation).

Figure 7.

A. fruticosa invading: A crop lands B grasslands C abandoned agricultural land west of Pecica locality D, E Mureș River riverbed F forest roads.

In relation to park zoning, out of the total mapped surface, 1.7% is located in the totally protected area (Felnac, Libus), 13.6% in the sustainable development area (largely in Pecica locality), while the remainder (84.7%) is in the sustainable management area, mainly spreading along the left bank of Mureş River.

The driving factors of A. fruticosa occurrence may vary from place to place. Many factors can affect the establishment and spread of invasive species (Underwood et al. 2004). They include the interaction of multiple environmental variables, such as elevation, precipitation and soil type, which constitute the species’ fundamental niche (Hutchinson 1957; Pysek et al. 2003). Invasive species have also been associated with areas of disturbance, either natural (e.g. fire or floods; Rejmánek 1989; Mack and D’Antonio 1998) or human related (Macdonald et al. 1988; Cowie and Werner 1993) and influenced by abiotic factors, such as historical land use and management (Mack et al. 2000).

In the present study, the selected explanatory variables encompass a significant share of the driving factors. Many studies have indicated that most of the analysed factors were also found to be important in other protected wetland areas in Romania (Dumitraşcu et al. 2011, Dumitraşcu et al. 2013, Kucsicsa et al. 2013; Dumitraşcu et al. 2014; Grigorescu et al. 2014), as well as in the other European countries (Blagojević et al. 2015; Radovanović et al. 2017; Delai et al. 2018). In the present study, the model shows that A. fruticosa is mainly controlled by the soil, groundwater’s availability, predominant vegetation type in the invaded areas and human-induced disturbance (forest fragmentation, roads extension). Specifically, amongst all analysed explanatory factors, erodisols, protisols and luvisols were found as the most significant categorical predictors for A. fruticosa occurrence in the Mureş Floodplain Natural Park. It should be noted that, in the study area, these classes of soil types cover about 80% of the total area, thus indicating a significant potential for A. fruticosa to spread in large areas of the park. The model also proved a significant relationship between the piezometric level and A. fruticosa distribution indicating species’ requirements for water availability in the soil.

A. fruticosa is considered a weak competitor in forests because it is usually excluded by tree species (Magyar 1960), but due to its fast growth, shading and probably its allelopathic effects (Elakowich and Wooten 1995) and nitrogen-fixing ability (Wang et al. 1999), it is a superior transformer in grasslands (Szigetvári 2002). In the present study, the model confirms the strong adaptability of A. fruticosa in grasslands areas. However, it can be noticed that the regression coefficients for highest values of NDVI also show the tendency of A. fruticosa to spread in the transitional woodland-scrub and afforested areas. This can indicate the adaptation capacity to any terrestrial ecosystems and, consequently, a tolerance for semi-shade. Furthermore, the model indicates the preference of A. fruticosa for fragmented forests, confirming that forest fragmentation could increase the ecosystems’ vulnerability to invasive species and habitat decline (Turner 1989).

It is widely known that roads can serve as corridors for the movement of invasive species (Christen and Matlack 2006), as well as for providing main habitats for establishment (Mortensen et al. 2009). In other Romanian protected areas, large areas covered by A. fruticosa were spotted in wetlands, along the forest roads (Dumitraşcu et al. 2013; Dumitraşcu et al. 2014). This can be explained by their influence in ecosystems fragmentation through creating suitable areas for invasive species growth. For the present-study, the regression results indicate that the proximity to roads was the predictor that has the most important contribution for A. fruticosa occurrence amongst the continuous driving factors. On the other hand, the continuity and the patterns of A. fruticosa occurrence along the roads show the very important role of roads in facilitating the movement of this invasive plant. Moreover, the road traffic close to the Park’s border can favour species’ expansion, having been known for its tolerance to polluting environments (Seo et al. 2008; Marian et al. 2010; Xiang et al. 2011). The development of the invasive species on contaminated areas was confirmed within another protected area of Romania (Comana Natural Park), where the large spread of A. fruticosa along the main roads and non-electrified railroad was observed (Dumitraşcu et al. 2011). The model does not demonstrate that A. fruticosa occurrence is significantly controlled by the proximity to settlements. However, the negative value of regression coefficients shows that the presence of the invasive plant species could be facilitated by human activities, indicating that the disturbed habitats inside and close to the settlements are easier to invade. This study also revealed that the proximity of aquatic surfaces has no significant influence on A. fruticosa occurrence. However, the positive regression coefficients, as well as the mapped areas with A. fruticosa (59%) within the first 500 m distance to Mureş River, could be explained through the favourable specific microclimate or fluvial processes within the riverbed which can favour the growth of invasive species to the detriment of riparian vegetation. In addition, the occurrence of A. fruticosa along the riverbed could be also explained by the fact that rivers are regularly considered natural vectors for invasive species dissemination (Fenesi et al. 2009).

In this respect, the authors consider that the more ecosystems and habitats are affected by disturbance, the more likely they become invaded by A. fruticosa. Thus, future forest fragmentation and clearing, the extension of the transportation network and the abandonment of the agricultural lands will increase the potential spread of A. fruticosa. Furthermore, planting A. fruticosa for different purposes (on the degraded lands, protection of dams or roads) will facilitate species’ invasion within the important habitats and ecosystems of the Park.

Invasive species may cause cascading effects in communities and/or affect both biotic and abiotic components of ecosystems (Charles and Dukes 2006) bringing in substantial costs to agriculture, forest and human health (Sîrbu et al. 2016b), as well as to ecosystem services, affecting ecosystem structure and function (Charles and Dukes 2006), loss of biodiversity or unique habitats. Invasive plants may decrease the suitability of soil for native species (Callaway and Ridenour 2004), disturbing soil formation, nutrient components or altering microbial communities. As a result, detailed knowledge of species’ ecological and geographic distribution is critical for effective conservation planning and modelling of its potential spread. However, data about most of species occurrence is sparse, resulting in incomplete information about species distribution, which leads to its difficult control and monitoring in sensitive areas (e.g. protected areas, wetlands). Hence, species distribution models attempt to provide detailed spatial data by relating presence of species to environmental predictors (Guisan and Thuiller 2005; Elith et al. 2006; Václavík and Meentemeyer 2009). In the current study, in order to identify and inventory A. fruticosa, as well as to develop a potential distribution model for the entire protected area, integrating GIS and logistic regression have been performed, given that we consider that the resultant map would provide the necessary information for the effective management of native ecosystems in the study area. Since the river systems are considered as main transport corridors for the invasive plants (Gallé et al. 1995), characterised by natural disturbances that create suitable sites for invasive species (Rood et al. 2010), it can be appreciated that A. fruticosa expansion could represent an important trans-border ecological issue given that Mureş River represents an important tributary of the Tisa River in the Danube Basin area. More than that, this can be critical given that the Danube is considered one of the most important routes for spreading invasive species, owing to its long distance, fluctuating water level and long-time anthropogenic presence, which facilitate these invasions (Pedashenko et al. 2012). Due to the resultant map indicating a significant susceptibility to invasion, we consider that, without careful management, the important habitats and essential ecosystem processes in large fluvial areas in the Mureş Floodplain Natural Park could be seriously disturbed in a relatively short time. Hence, the resultant outcomes could become important tools for the park’s administration to adopt appropriate planning strategies for the eradication/limitation of A. fruticosa in view of conserving the biodiversity of native flora and, finally, to provide the sustainable development of this protected area. On the other hand, the database containing the spatial distribution of A. fruticosa can serve the park’s administration and research as useful information about the location of the invasive species in order to monitor and carry out an assessment of the quantitative rates of dispersal in different habitats and ecosystems within the protected area, but also cross-boundary, knowing the relative continuity of the environmental conditions which provides suitable habitats for potential expansion. Furthermore, the results might be also used in other similar sites where A. fruticosa occurs in order to identify the areas that can be potentially invaded by this terrestrial invasive plant species. Given that this invasive plant is strongly associated with some landscape features such as soil type, depth to water, vegetation cover or roads, it is also essential to incorporate this knowledge into the assessment of potential spreading of other invasive terrestrial plant species in other sites. In addition, the probability map generated in this study can provide the basis for scenario analysis where independent variables can be improved and modified according to the specific biophysical and anthropogenic changes in an area.

Uncertainty is an inevitable component of invasion forecasts (Yemshanov et al. 2015) and modelling will always contain a level of errors resulting from a wide range of factors (Pearce et al. 2003), including insufficient sample size, measurements errors in the biological survey data or insufficient spatial resolution in the mapped environmental variables and impossibility or difficulty in integrating critical habitat variables and others factors (e.g. competition, dispersion). Hence, the resultant probability map should be used as a preliminary data on the potential distribution of A. fruticosa in order to identify regions with different probability and, consequently, to spot the areas that require more or less intensive monitoring of this invasive terrestrial plant species. Thus, important limitations and assumptions in the calibration of the model and generating the probability map have to be considered. The first is related to the A. fruticosa occurrence inventory dataset. In the present paper, we assume that all areas covered with A. fruticosa occurrence were not included in the analysis. However, several inconvenient factors (e.g. large extension of the area or inaccessibility in different sites) limited the mapping of all areas covered by A. fruticosa. We consider that the available datasets used to model the probability map were not sufficient to assess with the highest accuracy the potential occurrence of A. fruticosa. Thus, the estimated coefficients in the logistic regression models have associated estimation errors, the uncertainty decreasing by mapping more plant occurrence data and predictors (Horssen et al. 2002; Elith et al. 2006).

Another limitation refers to the unavailable datasets for the independent variables. Thus, the resulted pseudo R² values indicate that only 24.3% of A. fruticosa occurrence in the Mureş Floodplain Natural Park can be influenced by the analysed explanatory factors and the remaining percentage was influenced by other factors. Thus, in order to allow a better and realistic modelling of species’ spreading potential in the future at a much finer scale, more information on the spatial distribution of A. fruticosa combined with other predictors (e.g. soil nutrients and heavy metals content, past land-use changes, existing plant community) must be integrated. Moreover, the coarse resolution of the available soils and depth to water data have also restricted the accuracy of the model. One more limitation is related to the final probability map which does not reflect its temporal probability. As it is difficult or impossible to model seed dispersal at a regional scale (Goslee et al. 2006), the current results only display the spatial distribution potential of A. fruticosa depending on the analysed explanatory factors, without considering the seed dispersal vectors. In addition, environmentally suitable sites within the native distributional area may remain free from invasion because of biotic interactions, dispersal limitations or historical constraints (Ricklefs and Schluter 1993; Pulliam 2000). On the other hand, these restrictive factors may, at least in theory, differ or even be lacking in invaded areas (Jiménez-Valverde et al. 2011). As a consequence, it can be appreciated that, in the future, the spread might affect more or less other areas indicated by the current research.