*2.3. Methods*

We analyzed the spatial patterns associated with forest LULC change and related factors in an effort to seek efficient ways to achieve the SDGs for forests within the CCZ. Areas that underwent change from forested to non-forest landed (forest change areas) were extracted by superimposing the 2010 and 2016 land-cover maps. A GIS database was then established by spatial assessment along with related change factors such as topography (slope and elevation), accessibility (access to roads, buildings and the CCZ) and socioeconomic characteristics (cadastral value, local population and tourist attractions). This was then used for the forest change pattern and change factor analyses using Hot spot analysis and SEM, respectively (Figure 2). Both assessments used the basic geographic unit ri (village), "ri" is the smallest administrative division in South Korea.

**Figure 2.** Flow chart of the forest LULC pattern and change factor analyses.

#### 2.3.1. Forest LULC Change Spatial Analysis Using Spatial Statistics

In this study, spatial autocorrelation analysis (Hotspot analysis) was performed to analyze the pattern of forest land changing to other land.

LULC change is not independent of the surrounding environment, being influenced by adjacent land [20,31]. Therefore, we quantified interactions with surrounding areas using spatial autocorrelation, which measures the spatial similarity and/or dissimilarity of different regions in terms of proximity [32].

We used the Getis-Ord *Gi*\*, developed by [18], as the spatial autocorrelation index. This statistic calculates the proportions of the reference and neighboring spaces' values and measures the degree of spatial association between physical objects based on the concentration of weights. Getis-Ord *Gi*\* expresses spatial clusters of high and low values as hot and cold spots, respectively. When distribution patterns are randomly dispersed (without clusters) the values are close to zero [33].

$$Gi^\* = \frac{\sum\_{j=1}^n w\_{ij}\mathbf{x}\_i - \overline{\mathbf{X}}\sum\_{j=1}^n w\_{ij}}{s\sqrt{\frac{n\sum\_{j=1}^n w\_{ij}^2 - \left(\sum\_{j=1}^n w\_{ij}\right)^2}{n-1}}}$$

*xi* = attribute value of *i*

*wij* = spatial weight (spatial weight matrix value)

*s* = standard deviation

*n* = total number of cases

*i* and *i* are adjacent = 1

*i* and *i* are not adjacent = 0

#### 2.3.2. Forest LULC Change Factor Analysis Using SEM

#### SEM Background and Selection of Variables

SEM combines confirmatory factor analysis and path analysis, stemming from psychometrics and econometrics, respectively, through which causation and correlation between variables can be tested [34]. SEM is useful for estimating complicated causal relationships among multiple variables and simultaneously explaining direct and indirect effects between observable and unobservable ("latent") variables [35,36]. Furthermore, inter-variable relationships are displayed in schematic diagrams, allowing intuitive understanding of the analysis results [37]. In this study, SEM-based factor analysis involved multiple steps: selecting measurement variables with the potential to influence forest LULC, performing exploratory factor analysis to assess the inter-variable effects, building a study model based on the relationships among the measurement variables and analyzing the factors for forest LULC change by examining the causal relationships and associations among the variables.

SEMs have a structural component (path analysis) and a measurement component (factor analysis), in which the causal relationships are shown between dependent (endogenous) and independent (exogenous) variables and latent (constructs or factors) and observed (measured) variables, respectively. In the model we built for this study, the extent of forest LULC change per unit area was used as the measurement (dependent) variable. Latent variables including topographic (elevation (m) and slope (◦)), accessibility (m) (to roads, buildings and the CCZ) and socioeconomic (population distribution, tourist attractions and mean cadastral value (W)) factors were selected on the assumption that they could influence LULC change [12,38,39]. Elevation and slope information are constructed using digital topographic maps and accessibility is the distance (Unit: m) from the forest change areas of roads, buildings and CCZ. Population distribution and tourist attractions units are numbers and unit of mean cadastral value is won (W). The input data for these observable variables were calculated as averages by ri, except for tourist attractions, which used the total number. The units of the input variables were homogenized and the final variables selected by means of exploratory factor analysis [40].

#### SEM Construction and Assessment of Model Fit

Figure 3 presents the SEM for the forest LULC change area. The two left-hand columns show the dependent variables consisting of five measurement variables and one latent variable. The two right-hand columns show the independent variables consisting of three latent variables and eight measurement variables.

**Figure 3.** Framework of the SEM used in this study.

The SEM was evaluated using absolute and incremental fit indices (Table 2) to test how well the observed data fit into the model and how closely the measurement and latent variables were related to one another. An absolute fit index tests how well the proposed model predicts the sample covariance matrix and an incremental fit index measures the improvement in fit of the proposed model in comparison with the basic model [34,41].


**Table 2.** Fit indices and cutoff criteria used for SEM evaluation.

#### **3. Results and Discussion**

#### *3.1. Spatial Pattern Analysis of the Forest Change Area*

Statistically significant Hot spots for land-use change from forest to settlement in Ganghwa-gun, Gimpo-si, Yeoncheon-gun and Cheorwon-gun (Figure 4a). Jogang-ri in Gimpo-si and Naka-ri in Yeoncheon-gun displayed especially high Hot spot densities with Z-scores of 3.40 and 3.02, respectively. Frequent intense clusters of cold spots in Paju-si, Yeoncheon-gun and Cheorwon-gun also indicated the high rate of change from forest to settlement in these areas. Sung et al. [29] noted that deforestation in the CCZ is primarily due to the development of cropland and villages in Gyeonggi-do and the western part of Gangwon-do. Similarly, Lee [42] reported that the distribution density of villages in cities and counties with Hot spots accounted for about 74% of the overall distribution of villages in the CCZ. Village construction and expansion is especially intensive in Paju-si, Yeoncheon-gun and Cheorwon-gun, where agricultural activities are allowed on a residential or commuting basis.

Statistically significant Hot spots for land-use change from forest to cropland t in Gangnae-ri in Yeoncheon-gun, Woechon-ri in Cheorwon-gun and Mandae-ri in Yanggu-gun (Figure 4b). The latter displayed an especially high Hot spot density with a Z-score of 4.24. Hae'an-myeon and the surrounding area are characterized by intensive farming, with cropland accounting for 37.2% of the total area. Agricultural activities in this area are also diversifying into flowers, decorative plants, fruits and so forth, which continues to drive the conversion of forest into fields and orchards [43]. Park and Nam [44] reported on the continuously increasing land used for ginseng cultivation in this area since it hosted the "Gaeseong Ginseng Festival" in 2005. Sung et al. [29] also noted that expanding rice paddies and fields are a major cause of deforestation and emphasized that ginseng cultivation fields not only damage the primarily affected terrain but also cause secondary damage to surrounding terrain, weakening soils and degrading other vegetation. The increasing damage to forest edges in affected areas of Ganghwa-gun, Paju-si and Yeoncheon-gun can thus be attributed to the active ginseng cultivation operations in these areas in conjunction with broadly scattered small-scale agricultural land intermingled with forest.

Statistically significant Hot spots for land-use change from forest to grassland in Ganghwa-gun, Paju-si and Yeoncheon-gun (Figure 4c). The highest Hot spot densities by Z-score occurred in Cho-ri and Banjeong-ri in Paju-si and Majeon-ri and Dapgok-ri in Yeoncheon-gun (Figure 4c).

Statistically significant Hot spots for land-use change from forest to bare land in Ganghwa-gun, Yeoncheon-gun and Cheorwon-gun (Figure 4d). High-density Hot spots appeared in Gogu-ri in Ganghwa-gun and Hwoengsan-ri in Yeoncheon-gun (Z-scores: 3.94 and 3.59, respectively). Song et al. [45] reported that the LULC change rate was rapidly increasing after the governmen<sup>t</sup> announced its "Master Plan for Border Area Land Use Development" in 2011. "Master Plan for Border Area Land Use Development" is a plan announced by the Ministry of the Interior in 2011. This plan sets out the vision and goals for the development of CCZ. However, because this plan prioritizes development, forest researchers are worried that this plan will increase deforestation and degradation. Sung et al. [29] reported on continuous illegal clearing of trees and vegetation for soybean and ginseng cultivation in Yeoncheon-gun and Cheorwon-gun. The change from forest to bare land within the CCZ from Ganghwa-gun to Cheorwon-gun is thus ascribable to such land-use projects and illegal cropland development.

No Hot spots appeared in the spatial pattern indicating change from forest to wetland but cold spots appeared in Gangnae-ri in Yeoncheon-gun, Mandae-rin In Yanggu-gun and Seohwa-ri in Inje-gun (Figure 4e). Small-scale increases in wetland were at first assumed to be the result of precipitation but rainfall data for June 2010 and May 2016, the periods corresponding to the RS data used for the analysis, showed insufficient rainfall to cause significant changes in land cover during these periods [46,47]. A conversion of forest to wetland was interpreted as artefact, because no factual evidence was found in the field.

(**a**) **Figure 4.** *Cont.*

**Figure 4.** *Cont.*

**Figure 4.** Hot spot distribution maps of forest LULC change: (**a**) forest to settlement; (**b**) forest to cropland; (**c**) forest to grassland; (**d**) forest to bare land; (**e**) forest to wetland.

#### *3.2. Determination of Variables by Factor Analysis and SEM Fit Test*

Factor analysis is a statistical method used to test the validity of variables to be measured, in which variables with low validity are excluded from analysis or modified [48]. We used the Kaiser-Meyer-Olkin (KMO) test to assess correlations among variables for factor analysis and Bartlett's Test to determine the suitability of the factor analysis model. In general, a KMO value of 0.9 or above was deemed excellent, 0.8–0.9 good, 0.7–0.8 reasonable and 0.5 or less unacceptable [48]. The KMO and Bartlett's Test values for this study's variables were found to be good, at 0.815 and 0.000, respectively, demonstrating the suitability of the model used.

Commonality is the rate at which a specific variable is explained for various factor. Variables with a commonality of less than 0.4 are excluded from the factor analysis [49]. The variables used in this study were found to be suitable for factor analysis with communalities in excess of 0.4 (Table 3). All model fit indices were within acceptable ranges, so the model was judged to be suitable for hypothesis testing and quantifying causal relationships [37] (Table 4).


**Table 3.** Similarities of the variables extracted by factor analysis.


**Table 4.** Fit indices and cut-off criteria of the study model.

#### *3.3. Factor Analysis of Forest LULC Change*

The direct effect of the topographic factor on forest area change was −0.503, implying that topography had an inversely proportional effect on the extent of forest change. The greatest indirect effect was shown in the change to grassland (−0.428). Elevation was found to have a stronger impact (0.935) than slope (Figure 5a). Unlike grassland and bare land, the indirect effects of topography on changes to settlement, cropland and water were found to be minimal. Sung et al. [29] reported that increasing agricultural activity such as Ginseng cultivation is the main cause of vegetation deterioration and increasing bare land in the CCZ. Ganghwa-gun, Paju-si and Yeoncheon-gun were reported to be particularly affected by intensive agricultural activities. This is consistent with the factor analysis result that changes to grassland and bare land were more intense in the low-slope and low- elevation regions of the western CCZ.

The direct effect of the accessibility factor was −0.635, higher than the other two factors (topographic and socioeconomic) (Figure 5b). Accessibility had a negative effect on forest change, implying that change from forest cover to other land use increases in proportion with proximity to roads, buildings and the CCZ. The greatest indirect effect was shown in the change to bare land (−0.574), followed by grassland and cropland. The accessibility factor showed similar characteristics to the topographic factor, with higher impact on the change to grassland and lower impact on the change to bare land. This is consistent with the findings of Sung et al. [29] that deforestation and degradation increases in proportion to the level of forest land used for military facilities, strategic roads and villages.

The direct effect of the socioeconomic factor was 0.053, with forest change increasing with increases in tourist attraction distribution, local population and cadastral value (Figure 5c). The greatest indirect effect was shown in the change to grassland but the socioeconomic factor had very low effect on forest change overall, below 0.100 in both direct and indirect effects, presumably due to the intrinsically low numbers of tourist attractions and population in the CCZ.

(**a**) **Figure 5.** *Cont.*

(**c**) **Figure 5.** Factor analysis of the forest change area using the structural model: (**a**) topographic factor;(**b**)accessibilityfactor;(**c**)socioeconomicfactor.
