MODELLING FOREST COVER DYNAMICS IN ROMANIA (20012050) [309990]

MODELLING FOREST COVER DYNAMICS IN ROMANIA (2001–2050)

USING CLUE-S MODEL AND CORINE LAND COVER DATABASE

Abstract. The current paper focuses on future forest cover dynamics in Romania based on the data derived from CORINE Land Cover Database using the CLUE-S model (the Conversion of Land Use and its Effects at Small regional extent). The study aims to examine and analyse the various explanatory variables associated with the current forest pattern and to predict forest cover change (2001–2050) assuming the linear trend registered in the 1990–2000 period. The statistical validation was tested using ROC (Receiver Operating Characteristic) approach. Also, the model performance was assessed by comparing the predicted forest cover dynamics with the reference data (2000–2012). [anonimizat], climate, [anonimizat]. [anonimizat]. Generally, [anonimizat] 0.7%, especially in the Carpathian and Subcarpathian relief units. [anonimizat]/pastures to be abandoned and invaded by the transition vegetation. [anonimizat], grasslands and pastures in the most important agricultural areas of Romania. The CLUE-S model was applied for each Development Regions at the 1 ha resolution. Finally, simulated results were aggregated into single maps to analyse and understand the spatial distribution of the forest cover at national level.

Key words: [anonimizat]-S model, Romania.

1. INTRODUCTION

Land use and land cover changes and their impacts have become a [anonimizat]. [anonimizat]. [anonimizat]-environment conditions (Lambin and Geist 2006). Thus, [anonimizat] (Lin et al. 2007; Lin et al. 2009). Over the recent decades, a [anonimizat], possible future changes in land use and to understand the driving factors that control positive and negative feedback in land use change were developed (Lambin and Geist 2006). The change in the forest cover is of global concern since forest is an indispensable natural resource that provides not only a [anonimizat] a [anonimizat], climate change (Kumar et al. 2014). [anonimizat] (Kushwaha et al. 2011), degradation and fragmentation of habitats (Sun and Southworth 2013) and land degradation due to activation or intensification of present-day processes. Estimation of forest cover change is a major challenge. Studies related to forest cover change help in understanding phenomena like carbon dynamics, climate change and threats to biodiversity (Kumar et al. 2014).

Many studies using Geographic Information System (GIS) and various statistical analysis help to model and predict the dynamics of land use/land cover changes. Models are tools to support the analysis of the causes and consequences of land use changes in order to better understand the functioning of the land use system and to support land use planning and policy (Verburg et al. 2004). Spatially explicit models can provide information on the effects of certain policies and other driving factor changes and give an identification of regions and locations with probable high rates of land use change. The ‘hot-spots’ that are identified could represent priority areas for detailed analyses or policies (Verburg et al. 2004). As a result, one of the most used models to predict land use/cover changes based on driving factors is CLUE-S (Conversion of Land Use/Land Cover and its Effects at Small regional extent). This model is an example of advanced statistical land-use change (Veldkamp and Fresco 1996), an empirical multi-scale land-use change model developed for understanding and predicting the impact of biophysical and socio-economic forces that drive land-use change, used and validated in a wide range of applications (e.g. Veldkamp et al. 2001; Verburg and Veldkamp 2004; Castella et al. 2007; Wassenaar et al. 2007; Jinyan et al. 2007; Luo et al. 2010; Zhu et al. 2010). CLUE-S have been used in land change investigations across a wide range of scales of analysis and especially in several regions from Europe, Asia or Central America. In addition to simulation of multiple land use/cover changes, the model was applied to simulate forest dynamics and conservation (Wassenaar et al. 2007; Lin et al. 2009; Manuschevich and Beier 2016), to predict land-use changes and its impact on the groundwater or hydrological processes (Lin et al., 2007; Dams et al. 2008), to modelling urban growth (Li et al. 2014), to simulate agricultural land abandonment (Verburg and Overmars 2009; Renwick et al. 2013; Price et al. 2015) and to simulate extreme events (Chen et al. 2009; Zhou et al. 2013) or groundwater pollution related to land use/cover changes (Dams et al. 2008; Liu et al. 2014; Lima et al. 2015).

As from 1989, in Romania, the socio-economic and political changes caused by the fall of the communist regime have significantly influenced land use changes and altered forest management in region. The most important changes of that period were seen in the spatial dynamics of the main land use/land cover classes and their quality, a new type of landed property and land exploitation (Popovici 2008). Some of the negative effects of the land reform consisted in the excessive fragmentation of farmland, the noticeable degradation of production services in agriculture and land quality, deforestation, etc. (Bălteanu and Popovici 2010). Romania’s accession to the European Union and the implementation of the new Common Agricultural Policies represent a new evolution framework of land use and land cover pattern. Some changes in land use are related to this period (e.g. a reduced agricultural land fragmentation, a lower deforestation rate, intensive urbanisation). Over the past 26 years, forest land registered significant changes through deforestation or afforestation, especially in the hilly and mountain regions. The legal framework of forest management changed due to land privatization, reduction of state control, and changes in forest use regulations (Strimbu et al. 2005). The retrocession of forest land contributed to a large number of owners with small forest plots and to the uncertainty of maintaining these lands. Post-communist statistics report a shift in ownership to 30% private, 53% state and 17% other institutions (Munteanu et al. 2015). The transition period is characterized by unclear ownership regime, weak institutions, a low level of control and a poor development a protected areas network which led to an increased legal or illegal logging, especially on the privately-owned terrains, including inside protected area (Kuemmerle et al. 2009; Knorn et al. 2012; Popovici et al. 2013; Hanganu and Constantinescu 2015). The expansion of the forest-covered area is primarily due to natural regeneration, particularly in the mountainous and hilly regions. This process unfolded largely on deforested terrains, but also on abandoned farmlands, pastures which developed in the wake of declining shepherding. Artificial reforestation (e.g. planting after logging, calamities or establishment of a new forest on degraded lands) took place in small areas, directly depending on the financial sources (Dutca and Abrudan 2010; Popovici et al. 2013).

The present study aims to apply the CLUE-S model in order to describe and understand the process of the forest cover changes and estimate how they will change in the future in Romania under different explanatory factors (e.g. biophysical, socio-economic and accessibility). The outlined results were used as input data to evaluate the landslide hazard and risk in relation to the current and future land use/cover in Romania. This assessment was made within the recently completed project Disaster Risk Evaluation at a National Level (RO-RISK) under the Operational Programme for Administrative Capacity (POCA).

According to the current state of knowledge in Romania, the present paper represents a first attempt to simulate future forest cover dynamics at the national scale, modelled using CLUE-S methodology applied for each of the Development Regions of Romania. The resulted forest prediction maps can be useful tools for monitoring and quantifying spatial and temporal forest cover changes at national, and particularly at regional level. As a result, this could help identifying the most vulnerable areas in terms of forest losses, especially in the hilly and mountain regions, including within protected areas and thus, becoming helpful data in the sustainable forest management and forest policy. Furthermore, the outlined results could become important input data for the assessment of large-scale or small-scale natural hazards scenarios in relation to future forest cover changes.

2. MATERIALS AND METHODS

2.1. Study area

Romania is located in the South-Eastern part of Central Europe, at the contact with Eastern and Southern Europe (Fig. 1), at the crossroads of the main European access roads. It is a medium-sized European state covering a surface of 238,391 km2 and a population of 20,121,641 inhabitants (NIS Census data 2011), being the largest country in the south-east of the Continent (Bălteanu et al. 2016).

The moderate temperate-continental climate, the varied landforms, and the socio-economic conditions favour the great diversity of the land use/cover types with significant regional differences. Thus, the mountains and hills are dominantly covered by the forest and pastures/natural grasslands, while in the plains and tablelands arable lands prevail. In terms of structure, the agricultural land represents 61.2% of Romania’s area, the forest-covered areas 28.5%, the underwater terrains and ponds 3.5%, the built-up areas 3.1%, the roads and railways 1.6% and the degraded and unproductive grounds 2.1% (Romanian Statistical Yearbook 2012). According to the CORINE Land Cover database (2012), agricultural areas amounted to 57%, forest and semi-natural areas to 34%, artificial surfaces to 6%, and wetlands and water bodies to 3%. The ring-like display of the relief forms coupled with the climatic features, the diversity of soil resources, the vegetation cover and the human influence are key factors in the spatial distribution of the forest cover in Romania. Thus, from a total of almost 7,200,000 ha area covered by forest, 60% is located in the Carpathian Mountains, 35% in the hilly and tablelands units and only 5% in the plains and in the Danube Delta.

Figure 1. The study-area and the current forest-covered land. The Development Regions of Romania and the major landform units

Romania’s territory includes eight development regions (Fig. 1): North-West Development Region, Centre Development Region, North-East Development Region, South-East Development Region, South-Muntenia Development Region, Bucharest-Ilfov Development Region, South-West Oltenia Development Region and West Development Region. They act as regions without legal personality, which were intended in 1989 as territorial-statistical units for Eurostat and the absorption of European Structural Funds. Although conceived to be comparable territorial entities in surface-area, demographic size, social and economic features, yet some clear-cut regional disparities of global development between the less-developed South and East and the best-developed West and Central parts are only too obvious. The East and the South, which had a distinct rural character, experienced a post-1970 industrialisation; moreover, the South includes also Bucharest, Romania’s capital-city, a focal point of attraction. The central and southern parts benefitted from mature industries, better-developed services, and traditional urbanisation (Popescu and Săgeată 2016). The inter-regional disparities are also visible in forest cover areas and their dynamics. Statistically, in 2014, the largest forestlands occurred in the Centre and North-East Development Regions (over 1.2 mill. ha), West Development Region (1.2 mill. ha) and North-West Development Region (1 mill. ha), smallest forested surfaces have Bucharest-Ilfov Development Region (0.025 mill. ha), South-East Development Region (0.565 mill. ha) and South-Muntenia Development Region (0.671 mill. ha) (http://statistici.insse.ro/shop/).

2.2. Modelling future forest cover dynamics

2.2.1. Overview

The current study aims to analyse the future forest cover dynamics (2001–2050) and explain the relationship between this land use category and its explanatory factors. The results rely on the CLUE-S model applied to predict land use/cover categories dynamics in Romania.

Due to the particularities of the land use/cover changes at regional level, of availability of statistical data at the LAU2 levels, as well as the relative high spatial resolution used (1 ha), the model was applied for each Development Region of Romania – North–East (1), North–West (2), Central (3), South–East (4), South Muntenia (5), South–West Oltenia (6) and West (7). Finally, simulated results were aggregated into single maps to analyse and understand the spatial distribution of the future forest cover changes at national level. Generally, the synthetic illustration of the methodology is represented in figure 2.

Figure 2. The flowchart showing the methodology used to predict future forest cover dynamics

2.2.2. Data and data processing

Land use/cover data

Three datasets of the CORINE Land Cover Database (available at the European Environment Agency) were used in this study: years 1990, 2000 and 2012 (http://www.eea.europa.eu/publications/COR0-landcover, last accessed March 20th 2016). CORINE Land Cover database contains useful information for detecting land-use and land cover changes, for constructing a land cover account, as well as for modelling future land use and land cover changes. Based on GIS processing of the CORINE Land Cover data layer some studies at national, as well as regional level have been achieved (Feranec and Otahel 2001; Feranec et al. 2007, 2010, 2016; Kuemmerle et al. 2006, 2009; Otahel et al. 2002; Willems et al. 2005, Popovici et al. 2013; Hanganu and Constantinescu 2015, etc.). For our study, the CORINE Land Cover 1990 and 2000 datasets were used to assess past changes in land use/cover and to calculate past change trends. Therefore, for the start of the simulation (year 0), the 2000 datasets were used. The year 2012 was used to calculate the changes between 2000 and 2012, considered as reference map in order to evaluate the model accuracy (validation). According to CORINE Land Cover Database, eight land use/cover categories (dependent variables) were derived. These were generalised according to the level 3 of nomenclature classification (Table 1): artificial areas (class 0), arable lands (class 1), permanent crops (class 2), pastures and grasslands (class 3), scrub and/or herbaceous vegetation association (class 4), forests (class 5), open spaces with little or no vegetation (class 6) and heterogeneous agricultural areas (class 7).

Table 1. The generalized land use/cover classes according to CORINE Land Cover database

The explanatory factors

Identifying and selecting the potential driving factors of land cover/use changes are some of the most important steps of the current study. Mainly, this is based on the knowledge of the study area, on the results provided by previous researches, expert judgement, as well as on the availability of data. Thus, for the present study, the factors taken into account as explanatory were obtained from different sources, which included NIS, NMA, NRDIPAEP and ESRI Romania. Finally, depending on their availability, 31 driving factors were included in the model. Among these, forest cover dynamics was related to 24 explanatory factors: biophysical, socio-economic and accessibility (Table 2). Explanatory variables related to relief features (elevation and slope declivity) were derived from the Digital Elevation Model (30 m). The distance explanatory variables were calculated using a proximity analysis based on the minimum Euclidean distances (buffer = 1 km) to: the nearest towns, the nearest major roads, and the nearest major wood exploitation/processing centres. The socio-economic factors were derived from the available census data compatible with the requirements of the simulation. These driving factors and land use/cover classes for each raster cell were determined using ArcGis 10.1 and transformed into an ASCII text format (necessary for the simulation). Due to the different scale/resolutions of used spatial data, all spatial data were resampled at 100 x 100 m cell size for further analysis.

Table 2. List of the independent variables included in the logistic regression model

2.2.3. The configuration of the CLUE-S Model

The model is divided into a spatial and a non-spatial module. The non-spatial module calculates the demands for land use/cover based on analyses of explanatory factors. The spatial module translates these demands into land use/cover changes according to the probabilities and rules of different land use/cover types using a raster-based system (Verburg et al. 2002). In our simulations, the land-use conversion included the following aspects:

1. Land use/cover requirements (demand). The extrapolation of land use/cover change trends in the near future is a common technique to calculate the land use/cover requirements. In the current study, due to unavailability of a longer period of spatial data (before 1990) on land use/cover changes at country level, only one scenario was used. Thus, the authors calibrated the model in order to simulate land use/cover changes relying on the historical trend (HT). The HT is only based on the past changes (1990–2000) calculated according to the CORINE Land Cover Database.

According to the HT scenario, the linear trend was assumed for future land use/cover dynamics. For the forest cover, the annual change rate (1990–2000) and predicted areas under this scenarios are shown in Figure 3.

Figure 3. Past annual forest change rate (1990–2000) and linear trend extrapolated for the 2001–2050 period. Differences at the Development Regions of Romania

2. Conversion settings. The conversion settings for specific land use/cover types determine the temporal dynamics of the simulations (Verburg 2010). Therefore, two sets of parameters are needed to characterize the individual land use/cover types: conversion elasticity (ELAS) and land use/cover transition sequences (TS). The first parameter is related to the reversibility of land use/cover changes, the values ranging from 0 (easy conversion) to 1 (irreversible change). In this paper, ELAS was set according to the particularities of the study area: class 0 = 1; class 1 = 0.2; class 2 = 0.5; class 3 = 0.1; class 4 = 0.4; class 5 = 0.6; class 6 = 0.7; class 7 = 0.3. The TS (or conversion matrix) was configured based on the past changes and our understanding of the land use/cover system in the study area. Thus, all land use types can be converted into any land use category (TS = 1) with one exception, namely class 0 which cannot be converted into other categories (TS = 0).

3. Logistic regression analysis. The location characteristics determine the relative suitability of a location for the different land use/cover types. The CLUE-S model is using binomial logistic regression to define the empirical relationships between the land use/cover types and its explanatory factors. The CLUE-S model is able to effectively contain different types of driving factors and discriminate their contribution to land use changes by means of a logistic regression method, which can be expressed using Equation (1). Verburg et al. (2002) gave a more detailed description of the model.

(Eq. 1)

where Pi is the probability of a single grid cell for the occurrence of the considered land use/cover class i; X1, X2, …, Xn are the explanatory factors; β0, β1, …, βn are the beta values of the logistic regression for explanatory factors.

The evaluation of the inverse regression equation for each grid cell results in a probability layer that serves as input for the iterative land cover allocation for time step ti+1.

In the current study, the coefficients are estimated through logistic regression using the land use/cover categories (2000) as dependent variables and biophysical factors, socio-economic factors and accessibility as independent variables. The data were analysed by binary logistic using SPSS statistical software package (Statistical Package for the Social Sciences) to determine the relationship between each land use/cover type and the factors influencing them. The stepwise procedure of logistic regression was used (forward stepwise regression) to select the relevant explanatory factors from a larger set of factors that are assumed to influence or explain the land use/cover pattern. Thus, the variables that have no significant contribution to the explanation of the land use/cover pattern were excluded from the final regression equation. The regression coefficients are subsequently used to calculate the probability of a certain grid cell to be devoted to a certain land use/cover type given the value of driving factors in the year of the analysis. The response of these regression functions can then be visualized into raster probability maps of the study area based on the location suitability, given the probability of the occurrence of a certain land use/cover type per cell. The results of the logistic regression analysis were tested by using ROC (Relative Operating Characteristics), a measure of the goodness of fit of the logistic regression model (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 is a plot of the true positive rate (sensitivity) against the false positive rate (1–specificity) for different possible thresholds of the model. 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.

Finally, the simulation of land use/cover dynamics from 2001 to 2050 by setting rules of land transfer, demand, and other requirement parameters was carried out using Dyna-CLUE software.

2.2.4. Spatial data accuracy assessment

In order to evaluate the spatial goodness-of-fit of the simulation, a cross-classification map was created by overlaying the predicted forest dynamics map (2000–2012) resulted from the CLUE-S model and forest dynamics map (2000–2012) computed using the CORINE Land Cover database. Based on this, the six situations of agreement/disagreement were identified (Fig. 4): agreement for forest losses (the pixels representing deforestation in the predicted map overlap the pixels representing deforestation in the reference map); disagreement for forest losses (the pixels representing deforestation in the predicted map overlap the pixels representing forest persistence in the reference map); agreement for forest gains (the pixels representing afforestation in the predicted map overlap the pixels representing afforestation in the reference map); disagreement for forest gains (the pixels representing afforestation in the predicted map overlap the pixels representing deforestation in the reference map); agreement for forest persistence (the pixels representing forest persistence in the predicted map overlap the pixels representing forest persistence in the reference map); disagreement for forest persistence (the pixels representing forest persistence in the predicted map overlap the pixels representing forest losses in the reference map).

Figure 4. Agreement/disagreement identified between predicted and reference data. Bold line illustrate the situation with correct model simulations (agreement), while thin line illustrates errors of the simulation model (disagreement)

The measure of the agreement or accuracy between the predicted and reference data was assessed using the results of confusion matrix (error matrix), a common method used in remote sensing literature, also applied in land use modeling to compare the predicted data with real data (Geri et al. 2011; Verburg et al. 2002; Munroe et al. 2002; Liu et al. 2009; Ahmed et al. 2013; Kumar et al. 2014). In this study, the confusion matrix includes resulted values from the cross-classification map and displays the number of pixels correct and incorrect predicted for forest losses (class 1), forest gains (class 2) and forest persistence (class 3). In the matrix, simulated data is displayed as rows and the verification data (reference) as columns. The values on the diagonal are correctly predicted according to the reference data.

Table 3. The error matrix computed to evaluate the agreement/disagreement between the predicted and the reference map

Depending on the error matrix, different statistical measures of accuracy were calculated: Overall Accuracy (OA), Producer’s Accuracy (PA), User’s Accuracy (UA), and Kappa index. The OA was computed by dividing the total correct classified pixels by the total number of pixels in the error matrix. The PA was calculated by the total number of correct pixels in a category divided by the total number of pixels of that category. This indicates the probability of a reference pixel being correctly classified and is the measure of omission error. On the other hand, UA was calculated by the total number of correct pixels in a category divided by the total number of pixels that were classified in that category. Finally, the Kappa index (Cohen 1960) was determined by the Equation 2.

(Eq. 2)

where Po is the observed correct proportion (Overall Accuracy) and Pc is the expected correct proportion.

The expected correct proportion (Pc) is given by Equation 3:

(Eq. 3)

The results of Kappa coefficients were interpreted according to Landis and Koch (1977): ≤0 = poor accuracy; 0.01–0.20 = slight accuracy; 0.21–0.40 = fair accuracy; 0.41–0.60 = moderate accuracy; 0.61–0.80 = substantial accuracy; and 0.81–1 = almost perfect accuracy.

In order to appreciate the accuracy of the simulated data at national and regional level, these statistical measures was computed for the entire areas and for each Development Region of Romania.

3. RESULTS AND DISCUSSIONS

3.1. Spatial analysis of forest cover dynamics (1990-2000)

In order to predict the forest cover dynamics for 2001–2050, the past forest cover change, calculated for 1990–2000 period at Development Region level was assumed. Overall, the spatial analysis of land use/cover changes calculated based on CORINE Land Cover Database indicates that forest gains were higher than forest losses (Figure 5). Forest gains were identified as 90,214 ha, with an annual rate of net increase of forest area of 9,021 ha. Significant afforestation was noticed in the Carpathians Mountains (especially Eastern Carpathians). The analysis reveals that the main land use/cover categories which were converted into forests are class 4 (scrub and/or herbaceous vegetation association) and class 3 (pastures and grasslands), both classes representing almost 99% of total transition related to afforestation. The remaining percent (1%) was related to transition between class 1 (arable lands), class 7 (heterogeneous agricultural areas) and class 6 (open spaces with little or no vegetation) into forests in few areas of the Romanian Plain.

Figure 5. Changes in forest cover 1990 – 2000. The comparative chart for the major landform units of Romania

The afforested area lost accounted for almost 69,000 ha with an annual rate equal to 6,805 ha. The deforestation was highly concentrated in the mountain units from Eastern Carpathians and Apuseni Mountains. This flow was mainly related to the transition of the forest into class 4 and class 3, both classes representing almost 99% of total deforestation-related transition. The remaining percent (1%) was related to the transition of forests into class 1 and class 7 in some areas from the Moldavian Plateau, Transylvanian Tableland, and Romanian Plain.

3.2. The relationship between forest cover pattern and explanatory variables. Regression results

The results of the logistic regression models are showed in Table 4. These illustrate the relationship between forest and the explanatory variables related to β coefficients. According to the p values, all variables have statistical significance. Therefore, the results of logistic regression analysis, tested using ROC, show a good accuracy of the models. The AUC values (Table 4) indicate a good fit between the predicted forests cover data and the real (observed) forest cover data, with a prediction ability of 84.0% for Central and 91.3% for South-East Development Regions.

Table 4. Estimated β coefficients for the logistic regression models. Nagelkerke R Square and AUC values

** statistical significance at 99% (p < 0.01); all other values statistical significance at 99.9% (p < 0.001)

+ not significant correlation (B value = ~0.000) or removed by stepwise regression

3.2.1. Models interpretation

According to the spatial data used and the particularities of the study-area, the β coefficients show that the biophysical variables have the most important contribution to explaining the current spatial pattern of forest cover in Romania, with significant differences at the regional level. For categorical variables, regression coefficients for IV8 and IV5 suggest that luvisols and cambisols are ecological factors which have the most significant role for the forest cover suitability. Generally, for the other soil classes, regressions show a negative correlation, this suggesting the unsuitability for forests in areas which are favourable to arable lands or pastures and grasslands. Among the continuous variables, β coefficients for IV4 indicate suitability for forest cover according to the increase in the annual average of air temperature. In terms of topographic factors, the effect of the slope on forest cover is also evident, because of their low accessibility for logging, a higher slope indicating favourable areas for forest extension. On the other hand, the correlation between forest cover and elevation is not notable, the β coefficients ranking very low values, from 0.003 to 0.005. Also, the precipitations have beta values close to 0, showing small statistical relations to forest expansion. In terms of accessibility, the distance to the nearest major roads (IV22) and distance to nearest towns (IV24) show relative importance. The positive values indicate that the future forest growth is more likely to occur in areas that have low accessibility and connectivity with the nearest town which can represent important centres for wood processing. For each Development Region, the explanatory factors related to socio-economic indicators have no significant role in the current spatial pattern of the forest cover. Therefore, some of these explanatory factors were excluded from the regression models.

Overall, the values of Nagelkerke R Square indicator show that the forest cover is adequately described by the used explanatory factors. The values ranging from 0.41 to 0.56 indicate that all 24 independent variables together explain about 41–56% of the variance of dependent variable, with better values for West, South-East, and North-East Development Regions (Table 4). However, this percentage suggests also that several other biophysical and socio-economic explanatory factors might also have a significant contribution to forest cover dynamics (e.g. historical pattern of forests, distance to wood exploitation areas, distance to water courses, distance to forest edge, distance to sheepfolds, illegal logging, regional development programmes, hydrogeological conditions or other climate-related indicators).

3.3. Prediction of forest cover

3.3.1. Forest cover probability map

The main outcome of the logistic regression model is the probability map of the forest cover occurrence in relation to the analysed independent variables (Figure 6). The probability map shows that the forest occurrence will be higher in the current locations. Therefore, high probability values are located in areas currently covered by the herbaceous vegetation or transition vegetation (grass-woodland) developed close to the upper forest limits from the high mountain units from Southern and Eastern Carpathians or in the valleys and depressions from the Apuseni Mountains and Easter Carpathians. The lowest probability of forest occurrence remains the plains and tablelands, due to their suitability for agriculture.

Figure 6. The probability of forest cover occurrence, based on the logit model

3.3.2. Forest cover in 2050

At national level, the scenario modelled using CLUE-S methodology shows an increase in the forest-covered area with about 0.7% that is from approximately 6,985,000 ha to 7,035,000 ha, with the annual rate equal to 1000 ha.

Figure 7. Forest cover in 2000 (A) and 2050 (B). The identified changes (C) and comparative chart for the major landform units of Romania

At regional level, the scenario suggests an expansion of the existing forest-covered areas, especially in the depressions located in the Carpathian and Subcarpathian relief units (Fig. 7). In addition, in the mountain units from the Eastern and Southern Carpathians the rising of the upper forest is expected to have a positive trend, a phenomena also currently observed, mainly in relation to the declining of the grazing activities and/or the declaration of a large number of restricted protected areas (Fig. 8).

Figure 8. The tendency of rising the upper forest limits in the high mountains of the Romanian Carpathians (photo: G. Kucsicsa, 2016)

Therefore, the model shows that the expansion of forest will be mainly in relation to the decreasing in agricultural areas (Fig. 9), which can be attributed to the tendency of grassland/pastures, permanent crops or arable land to be abandonment and invaded by the transition vegetation, a phenomena observed especially in many hilly and plain regions of Romania (Fig. 10). The decrease in forest cover is expected to occur in the plains and tablelands units, especially in the Romanian Plain, the northern part of the Banat and Crișana Plain, Transylvanian Tableland and Moldavian Plateau. Also, significant deforestation in the northern part of the Apuseni Mountains and in few relief units from the Eastern and Southern Carpathians is also expected. The model shows that the forest losses will be largely related to the extension of arable lands, grasslands and pastures in the main agricultural areas from Romania or in the accessible mountain units, currently dominated by the heterogeneous agricultural areas or with intensive logging activities.

Figure 9. The dynamics of the land use/cover classes related to forest cover change in 2000 – 2050 (the transition map)

Figure 10. The tendency of the transitional vegetation to invade the agricultural mixed areas (A), grasslands (B) and vineyards (C) in the hilly and plain regions of Romania (photo: G. Kucsicsa, 2015)

That simulation results were part of the input data used to evaluate the landslide hazard and risk in relation to current and future land use/cover in Romania, investigated in the RO-RISK Project. In this respect, the landslide susceptibility map (General Inspectorate for Emergency Situations, 2016) in relation to the simulated forest cover dynamics show that in the entire study area is expected a sensitive degrease of susceptibility according to the forest cover expansion, but with significant regional difference. The spatial analysis explain that the increase of the landslides susceptibility is expected mainly within the Subcarpathians, the southern part of the Transylvanian Tableland and the central part of the Moldavian Plateau, areas currently the most affected by numerous landslides. This increase will be in relation to the conversion of the forest lands into pastures and grasslands, in areas with high and very high landslide susceptibility classes. Significant decrease of the landslides susceptibility is expected in the Getic Piedmont, Banat and Crișana Hills and in several mountain units from the southern part of the Eastern Carpathians (Curvature Carpathians) and Banat Mountains. The decrease will be associated with the extension of the forest cover to the detriment of pastures and grassland or heterogeneous agricultural areas.

3.3.3. Model accuracy

Accuracy assessment of the forest cover scenario was performed based on overlaying the predicted and the reference map. Figure 11 shows that the best results of agreement were obtained for the mountain regions were extended areas with forest losses or forest gains were predicted correctly (in northern and southern part of the Eastern Carpathians, western part of the Southern Carpathians, Banat Mountains). Therefore, significant deforested areas were accurately predicted for the Subcarpathians and Getic Piedmont, the central part of the Moldavian Tableland and some units from the western part of the Romanian Plain. Notable disagreement for forest losses and gains were obtained for Subcarpathians areas, Southern Carpathians and Apuseni Mountains.

Figure 11. Cross-classification map results of overlay process of predicted forest dynamic map (2000 – 2012) and reference map (real forest dynamics 2000 –2012)

Depending on the confusion matrix, different statistical measures such as Overall Accuracy (OA), User’s Accuracy (UA), Producer’s Accuracy (PA) and Kappa index were calculated for the entire study area and each Development Region (Table 5). OA equal with 85.03% for the entire study area and 82.14% – 88.01% for the regional level were obtained. The overall Kappa values for all the classification results are between 0.293 – 0.409, indicating a fair to moderate accuracy between predicted and real forest cover dynamics (2001–2012). This is the result of the high percentage of UA and PA for agreement in forest persistence (class 3) and low and medium percentage of the agreement in forest losses (class 1) and forest gains (class 2). Overall, the weak results of the Kappa index could be mainly caused by the:

– spatial data used in the simulation, namely the accuracy of the CORINE Land Cover Database; the different scales of the used explanatory factors; the availability of socio-economic statistical data, only at LAU2;

– unavailability of other relevant socio-economic or biophysical data related to forest dynamics;

– unavailability of other important dynamic driving factors (e.g. demographic or climatic scenarios);

– human-induced or natural unpredicted events (e.g. illegal logging, forests blow-downs, and forest fires);

– extrapolation of a short historical trend of land use/cover changes (only 10 years) to simulate a longer period (50 years);

– unexpected the political and institutional measures with influence on the land use/cover changes (Land Laws, land use planning, increasing the number of measures for afforestation projects, especially for degraded lands, Romania's accession to the European Union);

– environmental policies (declaration of a large number of protected areas, as of 2000, where the trend of deforestation rate has decreased, mainly inside the national parks rather than outside their boundaries; ecological reconstruction strategies and measures, especially within wetlands areas).

Table 5. The agreement statistic indicators for predicted data calculated for the entire study area and regional level

4. CONCLUSIONS

The current paper represents a study on the future forest cover dynamics (2001–2050) applying the CLUE-S model, based on the linear trend of the past changes calculated for 1990–2000 period. This assessment considered 24 biophysical, socio-economic and accessibility explanatory variables selected based on previous research findings, the knowledge of the study area, as well as on the availability of the spatial data. Thus, according to the used spatial data, the model shows that the biophysical variables have the most important contribution to explaining the current spatial pattern of forest cover in Romania. The highest positive β values suggest that soil types and the annual average of air temperature are factors which have the most significant contribution for forest cover suitability. Therefore, the regression coefficients indicate that the future forest growth is more likely to occur in areas that have low accessibility in terms of topography, communication network or nearest towns. On the other hand, the model shows that the explanatory factors related to socio-economic indicators haven’t played a significant role in the current spatial pattern of forest cover.

Overall, the resulted scenario reveals that forest cover has undergone continuous change, with significant differences at the level of major landform units of Romania. Consequently, for the entire study area an increase in the forest cover area with about 0.7% until the year 2050 is expected. At local level, the model suggests important spatial expansion of the existing forest areas in the depressions located in the Carpathians and Subcarpathians. Therefore, in the Southern Carpathians increase in the forest-covered area, especially in the highest areas where the current upper forest will have a positive trend is expected. The model shows that the expansion of forest will be especially in relation to the decreasing in agricultural areas, which can be attributed to the tendency of grassland/pastures or other agricultural lands to be abandoned. Significant deforestation will be largely related to the expansion of arable lands, grasslands and pastures in some of the most important agricultural areas in Romania (e.g. Romanian Plain, northern part of the Banat and Crișana Plain, Transylvanian Tableland and Moldavian Plateau) or in the mountain units with higher accessibility, currently dominated by the heterogeneous agricultural areas or with intensive logging activities.

In accordance with the results of the current simulation, used as input data for the national-level evaluation of the future landslide susceptibility in relation to land use/cover changes (RO-RISK Project), it can be concluded that significant increase of landslides susceptibility is expected, especially in the Subcarpathians, southern part of the Transylvanian Tableland and central part of the Moldavian Plateau, areas currently largely affected by a great number of landslides. In contrast, the decrease of the landslides susceptibility is expected in the Getic Piedmont, Banat and Crișana Hills and in several mountain units (especially from southern part of the Eastern Carpathians) triggered by the extension of the forest cover to the detriment of agricultural lands.

Further studies, including additional variables related to other biophysical and socio-economic explanatory factors, could better identify the effects of the explanatory variables on the forest conversion processes and could better simulate the future forest cover dynamics. The forest prediction maps could become useful tools to monitoring and quantifying spatial and temporal forest cover changes at national and regional levels. Thus, the prediction maps are more likely to turn into important input data to assess the magnitude of natural hazards in relation to the future forest cover changes.

Acknowledgements

The paper was elaborated in the framework of RO-RISK Project under the Administrative Capacity Operational Programme (Disaster Risk Assessment at national level – SIPOCA 30) and the project “Atlas of Environment” – fundamental study made under the research plan of the Romanian Academy’s Institute of Geography.

This study was made possible with the support of the full version of CLUE soft available at the IVM Institute for Environmental Studies (http://www.ivm.vu.nl/en/Organisation/departments/spatial-analysis-decision-support/Clue). Also, we would like to thank the European Environment Agency for the CORINE Land Cover database.

References

Ahmed B, Ahmed R, Zhu X (2013) Evaluation of Model Validation Techniques in Land Cover Dynamics, SPRS Int. J. Geo-Inf 2:577-597

Bălteanu D, Dumitrașcu M, Geacu S, Mitrică B, Sima M (eds.) (2016) Romania. Nature and Society, Romanian Academy Publishing House, 690 p., in Romanian

Castella J-C, Pheng Kam S, Dinh Quang D, Verburg PH, Thai Hoanh C (2007) Combining top-down and bottom-up modelling approaches of land use/cover change to support public policies: application to sustainable management of natural resources in northern Vietnam, Land Use Policy, 24:531–545

Chen Y, Xu Y, Yin Y (2009) Impacts of land use change scenarios on storm-runoff generation in Xitiaoxi basin, China, Quaternary International, 208 (1): 121–128

Cimini Dora, Tomao A, Mattioli W, Barbati A, Corona P (2013) Assessing impact of forest cover change dynamics on high nature value farmland in Mediterranean mountain landscape, Annals of Silvicultural Research, 37 (1): 29-37

Cohen J (1960) A coefficient of agreement for nominal scale, Educational and Psychological Measurement, 20, 37–46

Dams J, Woldeamlak ST, Batelaan O (2008) Predicting land-use change and its impact on the groundwater system of the Kleine Nete catchment, Belgium, Hydrol. Earth Syst. Sci., 12 (6):1369-1385

Dutca I, Abrudan I V (2010) Estimation of Forest Land-Cover Change in Romania, between 1990 and 2006, Bulletin of the Transilvania University of Brasov, 3 (52):33–36

Feranec J, Hazeu G, Christensen S, Jaffrain G (2007) CORINE land cover change detection in Europe (case studies of the Netherlands and Slovakia), Land Use Policy, 24:234–247

Feranec J, Jaffrain G, Soukup T, Hazeu G (2010) Determining changes and flows in European landscapes 1990–2000 using CORINE land cover data, Applied Geography, 30:19–35

Feranec J, Otahel J (2001) Land cover of Slovakia, Veda, Bratislava, 124

Feranec J, Soukup T, Hazeu G, Jaffrain G (2016) European Landscape Dynamics: CORINE Land Cover Data. CRC Press, Boca Raton, USA, 337

General Inspectorate for Emergency Situations (2016) Country report 5.1 Conditionality Romania, National Risk Assessment – RO RISK project (co-financed under EFS through the Operational Program Administrative Capacity 2014 – 2020), https://www.igsu.ro/documente/RO-RISK/Raport_Final_de_tara.pdf

Geri F, Amici V, Rocchini D (2011) Spatially-based accuracy assessment of forestation prediction in a complex Mediterranean landscape, Applied Geography, 31:881–890

Hanganu J, Constantinescu A (2015) Land cover changes in Romania based on Corine Land Cover inventory 1990–2012, Rev. Roum. Géogr./Rom. Journ. Geogr., 59 (2), 111–116

European Environment Agency, http://www.eea.europa.eu/publications/COR0-landcover, last accessed March 20th 2016

Jinyan Z, Xiangzheng D, Qun’ou J, Nana S (2007) The application of system dynamics and CLUE-S model in land use change dynamic simulation: a case study in Taips County, Inner Mongolia of China, Proceedings of the Conference on System Science, Management Science & System Dynamics, 2781–2790

Knorn J, Kuemmerle T, Radeloff VC, Szabo A, Mindrescu M, Keeton WS, Abrudan I, Griffiths P, Gancz V, Hostert P (2012) Forest restitution and protected area effectiveness in post-socialist Romania, Biological Conservation, 146 (1): 204–212

Kuemmerle T, Hostert P, Radeloff VC, Perzanowski K, Kruhlov I (2007) Post-Socialist Forest Disturbance in the Carpathian Border Region of Poland, Slovakia, and Ukraine, Ecological Applications, 17(5): 1279–1295

Kuemmerle T, Müller D, Griffiths P, Rusu M (2009) Land use change in Southern Romania after the collapse of socialism, Reg Environ Change, 9:1–12

Kuemmerle T, Radeloff VC, Perzanowski K, Hostert P (2006) Crossborder Comparison of Land Cover and Landscape Pattern in Eastern Europe Using a Hybrid Classification Technique, Remote Sensing of Environment, 103: 449–464

Kumar R, Nandy S, Agarwal R, Kushwaha SPS (2014) Forest cover dynamics analysis and prediction modeling using logistic regression model, Ecological Indicators, 45: 444–455

Kushwaha SPS, Nandy S, Ahmad M, Agarwal R (2011) Forest ecosystem dynamics assessment and predictive modelling in eastern Himalaya, International Society for Photogrammetry and Remote Sensing Archives, 38, Workshop Proceedings: Earth Observation for Terrestrial Ecosystems, 8 November, Bhopal, India

Landis J, Koch G (1977) The measurement of observer agreement for categorical data, Biometrics, 33: 159–17

Li W, Wu C, Zang S (2014) Modeling urban land use conversion of Daqing City, China: A comparative analysis of ‘‘top-down’’ and ‘‘bottom-up’’ approaches, Stoch Environ Res Risk Assess, 28 (4): 817–828

Lima ML, Romanelli A, Massone HE (2015) Assessing groundwater pollution hazard changes under different socio-economic and environmental scenarios in an agricultural watershed, Science of the Total Environment, 530: 333-346

Lin PY, Nien-Ming Hong NM, Pei-Jung Wu PJ, Lin, CJ (2007) Modeling and assessing land-use and hydrological processes to future land-use and climate change scenarios in watershed land-use planning, Environ. Geol., 53:623–634

Lin YP, Verburg PH, Chang CR, Chen HY, Chen MH (2009) Developing and comparing optimal and empirical land-use models for the development of an urbanized watershed forest in Taiwan. Landscape and Urban Planning, 92 (3): 242–254

Liu M, Hu Y, Chang Y, He X, Zhang W (2009) Land Use and Land Cover Change Analysis and Prediction in the Upper Reaches of the Minjiang River, China, Environmental Management, 43:899–907

Liu M, Li C, Hu Y, Sun F, Xu Y, Chen T (2014) Combining CLUE-S and SWAT models to forecast land use change and non-point source pollution impact at a watershed scale in Liaoning Province, China, Chinese Geographical Science, 24 (5): 540–550

Luo G, Yin C, Chen X, Xu W, Lu L (2010) Combining system dynamic model and CLUE-S model to improve land use scenario analyses at regional scale: A case study of Sangong watershed in Xinjiang, China, Ecological Complexity, 7 (2):198–207

Manuschevich D, Beier CM (2016) Simulating land use changes under alternative policy scenarios for conservation of native forests in south-central Chile, Land Use Policy, 51:350–362

Mas J-F, Filho BS, Pontius Jr RG, Gutiérrez MF, Rodrigues H (2013) A Suite of Tools for ROC Analysis of Spatial Models, ISPRS Int. J. Geo-Inf., 2:869–887

Munroe DK, Southworth J, Tucker CM (2002) The dynamics of land-cover change in western Honduras: exploring spatial and temporal complexity, Agricultural Economics, 27:355-369

Munteanu C, Niță MD, Abrudan IV, Radeloff CV (2015) Historical forest management in Romania is imposing strong legacies on contemporary forests and their management, Forest Ecology and Management 361 (2016):179–193

National Institute of Statistics, Population and Dwellings Census 2011, Bucharest, http://www.recensamantromania.ro/en/

National Institute of Statistics. Romanian Statistical Yearbook 2012. http://www.insse.ro/cms/files/Anuar%20arhive/serii%20de%20date/2012/pdf/ASR%202012_rom.pdf

National Institute of Statistics. TEMPO-online database, http://statistici.insse.ro/shop/

Otahel J, Feranec J, Suri M, Cebecauer T, (2002) Land cover of Slovakia: assessment of its changes, In: Bicík I et al. (eds.): Land Use/Land Cover Changes in the Period of Globalization. Proceedings of the IGU-LUCC International Conference, Prague, 2001. KSGRR PrF UK, Prague, 100–105

Ping Zhang P, Liua Y, Panb, Y, Yua, Z (2013) Land use pattern optimization based on CLUE-S and SWAT models for agricultural non-point source pollution control, Mathematical and Computer Modelling, 58: 588–595

Popescu C, Săgeată R (2016) Regiunile de dezvoltare și politica de dezvoltarea regională, în vol. România. Natură și Societate, Edit. Academiei Române, București, pp. 609-612

Popovici E-A, Bălteanu D, Kucsicsa Gh (2013) Assessment of changes in Land-Use and Land-Cover pattern in Romania using Corine Land Cover database, Carpathian Journal of Earth and Environmental Sciences, 8 (4): 195–208

Price B, Kienast F, Seidl I, Ginzler C, Verburg PH, Bolliger J (2015) Future landscapes of Switzerland: Risk areas for urbanisation and land abandonment, Applied Geography, 57:32–41

Renwick A, Jansson T, Verburg PH, Revoredo-Giha C, Britz W, Gocht A, McCracken D (2013) Policy reform and agricultural land abandonment in the EU, Land Use Policy, 30 (1): 446–457

Story M, Congalton R (1986) Accuracy assessment: A user’s perspective, Photogramm. Eng. Remote Sens. 52 (3):397–399

Strimbu BM, Hickey GM, Strimbu VG (2005) Forest conditions and management under rapid legislation change in Romania. The Forestry Chronicle, 81(3):350–358

Sun J, Southworth J (2013) Remote sensing-based fractal analysis and scale dependence associated with forest fragmentation in an Amazon tri-national frontier, Remote Sens. 5:454–472

Veldkamp A, Fresco LO (1996) CLUE: a conceptual model to study the conversion of land use and its effects, Ecological Modeling, 85:253–270

Veldkamp A, Verburg PH, Kok K, Koning GHJ, Priess J, Bergsma AR (2001) The need for scale sensitive approaches in spatially explicit land use change modelling, Environmental Modeling and Assessment, 6:111–121.

Verburg PH (2010) The CLUE Modeling Framework – course material, Institute for Environmental Studies, University Amsterdam, 53 p

Verburg PH, Overmars KP (2009) Combining top-down and bottom-up dynamics in land use modeling: exploring the future of abandoned farmlands in Europe with the Dyna-CLUE model, Landscape ecology, 24 (9):1167–1181

Verburg PH, Schot PP, Dijst MJ, Veldkamp A (2004) Land use change modelling: current practice and research priorities, GeoJournal, 61:309–324

Verburg PH, Soepboer W, Limpiada R, Espaldon MVO, Sharifa M, Veldkamp A (2002) Land use change modelling at the regional scale: the CLUE-S model. Environmental Management 30: 391–405

Verburg PH, Veldkamp A (2004) Projecting land use transitions at forest fringes in the Philippines at two spatial scales. Landscape Ecology, 19:77–98

Wassenaar T, Gerber P, Verburg PH, Rosales M, Ibrahim M, Steinfeld H (2007) Projecting land use changes in the Neotropics: The geography of pasture expansion into forest. Global Environmental Change, 17 (1): 86–104

Weber JL (2007) Implementation of land and ecosystem accounts at the European environment agency. Ecological Economics, 61(4): 695–707

Wells MP, Williams MD (1998) Russia’s Protected Areas in Transition: The Impacts of Perestroika, Economic Reform and the Move Towards Democracy, Ambio, 27(3): 198–206

Willems E, Lemments T, Buffaria B (2005) Utilisation of CLC 90&2000 data for monitoring the impact of CAP developments on the rural landscape, In: Trends of some agro-environmental indicators in the European Union (Report EUR 21565):15–349

Zhou F, Xu Y, Chen Y, Xu CY, Gao Y, Du J (2013) Hydrological response to urbanization at different spatio-temporal scales simulated by coupling of CLUE-S and the SWAT model in the Yangtze River Delta region, Journal of Hydrology, 485:113–125

Zhu Z, Liu L, Chen Z, Zhang J, Verburg PH (2010) Land-use change simulation and assessment of driving factors in the loess hilly region: a case study as Pengyang County, Environmental Monitoring and Assessment, 164 (1–4):133–142