Finding Sprawl Factors and Pirate Development: Based on Spatial Analysis of Population Grid Changes from 2014 to 2022 in SMA, South Korea

Abstract

This comprehensive study explores urban sprawl in the Seoul Metropolitan Area (SMA), emphasizing its rising intensity and complexity despite previous public-led planning efforts. The study aims to visualize the spatial patterns of sprawl and identify influencing factors through spatial regression analysis using grid-based population data created from actual population distributions. This approach fills a gap in the existing literature by moving beyond administrative-level analyses prone to ecological fallacies. This study scrutinizes the dynamics of population change in Seoul Metropolitan Areas (SMAs) in Korea over a decade, focusing on the predatory aspect of urban sprawl. Using grid-based population data and spatial regression analysis, the study finds that population growth is concentrated in unplanned areas with high development benefits. Three key hypotheses were examined: (1) Areas with high development potential, measured through factors like land prices and development plans, attract predatory development; (2) Improved transportation infrastructure encourages population inflow; (3) Non-urban land use, especially bare land, attracts population growth. The results offer important policy implications, particularly for preparing areas with low land prices and improving transportation infrastructures for future population influxes. Monitoring is particularly crucial in areas where development plans are already in place or where there is a high percentage of bare land.

1. Introduction

1.1. Background

Urban growth has been accompanied by sprawl since before the advent of the automobile (Hall, 2014), and this has generally been perceived as an inefficient land use phenomenon [1,2]. Along with suburbanization, the control of sprawl has become one of the main challenges of contemporary urban planning. In the 20th century, modern urban planners have proposed various alternatives to achieve this task: greenbelts, development permit systems, subdivision regulations, and even direct control of population movements have been conceived and used. On the other hand, policies intended to address other urban problems have often resulted in sprawl, such as subsidies and financial programs for housing, and deregulation to counteract rapid out-migration. Sometimes, policies intended to prevent sprawl, such as transit subsidies, have resulted in sprawl [3,4].

While definitions of urban sprawl vary, it can generally be understood as the spread-out and discontinuous character of low-density urban space in terms of the amount of land use, landscape, and density of people [5,6]. Originally, sprawl referred to the spread-out, non-contiguous, low-density, skipped-over development on the outskirts of cities that emerged in the United States in response to increased automobile use [7,8]. However, it is not unique to the United States; the same phenomenon has occurred in Europe, Latin America, and Asia as automobile dependence has increased [8,9,10].

Sprawl itself is not an entirely negative phenomenon, as it exists in response to the demands of residents and local governments. However, at the very least, it has side effects such as the decline of urban centers, increased use of private automobiles and the resulting overconsumption of energy, traffic congestion, air pollution, and natural degradation of urban neighborhoods [11,12]. Furthermore, even if sprawl appears disorganized, it may not always result from a lack of planning. Partial planning can lead to disorder in the organic whole of a city [13].

The Seoul Metropolitan Area (SMA), the largest megacity region in South Korea, has been reported to have experienced urban sprawl since the 1990s [14]. However, SMA is also believed to have experienced little sprawl relative to its size [15]. Rather, it has been argued that a polycentric and orderly network of satellite cities has been formed through the public-led supply of new towns [16,17]. By monopolizing the supply of large-scale new towns to public institutions, planned and compact space utilization was achieved. At the same time, transportation infrastructure was provided to facilitate commuting, schooling, and leisure traffic for a large population of more than 25 million people [18]. In addition, the preference of the Korean middle and upper classes for high-rise apartment buildings became a driving force for population concentration in planned new towns [19,20].

However, in recent years, there has been growing concern that the SMA is experiencing significant sprawl. Since 2017, the population outside the SMA has been growing more rapidly due to a soaring housing price surge inside Seoul City, which is the core of SMA. This phenomenon was exacerbated more by the COVID-19 outbreak in South Korea in 2020. Meanwhile, new towns are being developed sprawlingly within the SMA despite the principled intention of planned supply. The public transportation network connecting them is also becoming overly complex and inefficient. This phenomenon is because the supply of new towns has been on an ad hoc basis as the population grows, and the subsequent supply of new towns has been influenced by the distribution of land prices and transportation networks in existing new towns [21,22].

1.2. Research Purpose and Literature

This study aims to estimate the factors affecting the spatial distribution pattern of the SMA population at an empirical level. Specifically, this study compares the population grid data of SMA in 2014 and 2023 to analyze the sprawl pattern over the past decade. In South Korea, grid-based population data have recently begun to be produced and utilized. It is based on the addresses of the population, allowing for clear visualization and useful spatial regression analysis. Since the late 2010s, Statistics Korea (National Statistics Office) and private companies have created and released population grids based on actual census data. This study unfolds as follows:

Define SMAs operationally and introduce the context of development within them.

Visualize the population grid and other spatial information using GIS software, and build the hypotheses.

Introduce spatial regression analysis techniques, and construct variables for testing the hypotheses.

Conduct the analysis to verify and discuss the results.

While many studies measure sprawl and confirm its existence [23], there have been few studies of sprawl based on population grids, which we present here. It is because, despite the logical connection between sprawl and population, in many countries, population has been aggregated only at the administrative or statistical district level [17,24,25,26,27]. Of course, some studies have used population grids, but they are often geometrically estimated based on administrative area data rather than actual censuses [28,29]. In contrast, with the development of GIS technology, many studies have converted the area of urbanized land or the density distribution of buildings into grids [23,30].

Studies that have attempted to identify the driving factors of sprawl have either analyzed city-unit samples or, even when dealing with internal urban space, have focused on the spread of built area rather than population distribution. Deng et al. (2020) [31] examined urban sprawl trends in cities across China from 1992 to 2015 and argued that population growth is the most important factor, but topographic characteristics such as flat land area also play an important role. Zhang and Pan (2021) [32] examined the factors of sprawl in 343 cities in China and concluded that the gross regional product of tertiary industries and real estate development investment were important factors. Sprawl is measured by comparing the population growth rate and the built-up area growth rate of each city, which does not take into account the spatial distribution of the population within the city.

Seevarethanm et al. (2022) [33] considered sprawl as the spread of built areas and conducted a logistic regression analysis incorporating socioeconomic and topographical factors. In particular, the stricter the policy regulation, the less sprawl was found, and sprawl was more prevalent in areas with low land value. In addition, transportation and accessibility to the city center were covered, and a comprehensive analysis was conducted, but sprawl was not covered from the perspective of population distribution. Hosseni and Hajilou (2018) [34] did not directly address the spatial aspect but extracted 22 factors of urban sprawl based on more than 60 previous studies. Then, through an expert survey, they identified population growth, land value, political divisions, and transportation policies as important factors related to urban sprawl in Iran.

Therefore, quantitative studies that measure sprawl and analyze its drivers based on population grids are even rarer. Analyzing population change at the administrative district level is prone to ecological fallacies, as it is challenging to identify the spatial factors that influence population change [35]. Analyzing only land cover changes has the limitation that it does not reflect the population, which is the core of sprawl [29,36].

However, what existing studies have in common is that population growth, land prices, transportation accessibility, ease of development, and strictness of policy regulations play an important role in the spread of sprawl. This study aims to compensate for the weaknesses in the literature by analyzing the spatial dimension of factors such as land value level, change in road ratio, availability of developable land cover, and regulated land use by converting them to a grid scale.

2. Data and Hypothesis

2.1. Study Area and Its Developmental Contexts

The study area is the SMA, but in the empirical regression analysis, we exclude island areas and the border areas between North and South Korea due to data restrictions. In general, the definition of the SMA consists of Seoul City, Incheon City, and Gyeonggi Province. In recent years, it has been argued that the SMA has expanded beyond Gyeonggi Province’s boundaries, but it has traditionally firm boundaries [17,37] (Figure 1).

The total population of the SMA in 2023 is about 26.12 million, accounting for 50.5% of the total population of South Korea, and the GRDP in 2021 is about KRW 1075 trillion, accounting for 52.7% of the GDP. The population of the SMA has shown continuous and rapid growth, increasing by 2.5% annually for 70 years from a population of 4 million in the 1950s to the present. By contrast, the population of Seoul City, the core of the SMA, reached 10 million in the 1990s and has since declined to 9.66 million (Figure 2).

According to the Korean government’s land cover classification data, the urban area, which accounted for 4.3% of the SMA in 1985, increased to 9.0% in 2015. In particular, the urban area of the “SMA excluding Seoul” increased 2.7 times, from 301 km² in 1985 to 809 km² in 2015.

Large-scale residential developments in Seoul were generally built inside the city’s boundaries until the 1980s. From the 1990s, the greenbelt adjacent to the city was gradually removed. The city’s administrative area doubled in size in the 1960s, while greenbelts were designated on its outskirts. Greenbelt designations make it nearly impossible for private or public development.

However, in the 1980s, as demand for housing in the Seoul area skyrocketed, a loophole was created to utilize the greenbelt with the enactment of the Land Development Promotion Act, or the so-called New Towns Act, which allowed public agencies to expropriate affordable greenbelt land en masse to provide large-scale housing. In fact, since the 1990s, new towns have been developed within the greenbelt and similar development-restricted areas to accommodate the population that left the provinces for Seoul in search of jobs.

In principle, land designated as a greenbelt cannot be developed except by the government and public institutions, so government agencies that develop new towns could acquire land cheaply. In addition, by connecting the new towns to Seoul with urban highways and urban railways while maintaining development regulations on the land around the new towns, the government was able to sell land in the new towns at a high price to cover not only the cost of internal infrastructure but also the cost of building a wide-area transportation system. For example, Bundang New Town initially accepted land that had taken more than an hour to commute to Seoul but built a mass transit system that improved the distance to 30 min and sold the land at a high price. This business model continued to be utilized throughout the early 2000s and into the late 2010s. This has resulted in the spread of population centers around transportation infrastructure across the SMA (Figure 3).

As a business model, the expropriation of heavily restricted land and the provision of new mass transit has been very effective, but as this approach has been repeated, the supply of new towns has been affected by the existing ones. The first new towns in the 1990s were developed within 30 min of Seoul by mass transit. However, the new towns in the early 2000s were developed much further away, an hour outside the development restriction zone, and decided to extend the existing transportation system. Even then, complaints increased as transportation development was delayed. In the 2010s, new towns were located on existing undeveloped affordable land. The new towns of the 2010s were located on existing undeveloped low-cost land. In order to quickly utilize the land and supply affordable housing, the restricted development zones were lifted to develop the land, and the new towns of the 2010s became sandwiched between the new towns of the 1990s and 2000s.

Constructing new towns and transportation infrastructure simultaneously has formed an interconnected transportation network in the entire metropolitan area, creating conditions for private capital to promote sprawl. As shown in Figure 4, if all development outside the existing city in the CBD-A section is regulated and a public new town is developed in the B-S section, land prices in the A-B section are kept at a low level. However, if development regulations are loosened, section A-B can become the best prey for private capital, which can reliably earn maximum profits. It is where predatory development occurs, and sprawl spreads.

We can distinguish between traditional sprawl around urban centers; and infill or in-between sprawl, where low-density development occurs between existing, planned, high-density neighborhoods. It is afraid that the weakening of development regulations with the maturation of urban areas and the availability of transportation networks may lead to urban sprawl in the form of infill (Figure 4).

2.2. Data Sources

First, the data that will directly represent the sprawl phenomenon are the population grid of the SMA. We will compare the population grid in October 2014 and April 2023, and the amount of change is the dependent variable. Statistics Korea provides population grids at various scales based on the resident register. In principle, all Koreans and foreigners living in Korea for more than six months register their addresses in the resident registration system, which Statistics Korea can use. The most miniature scale is 100 m × 100 m, and the largest is 100 km × 100 km. This study uses a 1 km × 1 km population grid. It is a reasonable scale because even a new town, which is a large population center, is usually 1–2 km on a side. Larger scales are challenging to analyze, and more minor scales have technical limitations such as extensive computations and are not expected to improve the results significantly.

Additional data researched included published land prices, land use cover, road networks, highway access, planned developments, greenbelts, and businesses in the SMA. Land prices and highway accessibility are gridded data from sources such as the population grid. In South Korea, all land is assigned an officially certified land value and used as a tax base. We calculated this as an average value per grid. Other data were compiled from various government ministries. In particular, the road naming data used for the road network can be used to identify newly constructed roads since 2014. The actual width of roads on the numerical map was used in conjunction with this data. To consider the effect of corporate location, we used geocoded business addresses from SMTp, a big data database of companies in the private sector, which includes all companies that prepare financial reports (

Table 1. Data sources.

Data	Description	Source
Population	Population Grid (1 km × 1 km) in 2014, 2023	Korea National Geographic Information Institute
Land Price	Land Price Grid (1 km × 1 km) in 2014
Land Use Cover	Land Use Cover Map in 2010	Ministry of Environment, Korea
Road Network	Road Name Enlist Shape in 2022	Ministry of Land, Infrastructure and Transportation, Korea
Road Cover	National Numerical Map in 2022	Korea National Geographic Information Institute
Highway Accessibility	Distance Grid (500 m × 500 m) to the nearest Highway entrance in 2016	Korea National Geographic Information Institute
Development Plan constructed area	Development Plan constructed Area Shape in 2023	Ministry of Land, Infrastructure and Transportation, Korea
Greenbelt	Restricted Development Area Shape in 2023	Ministry of Land, Infrastructure and Transportation, Korea
Enterprise Locations	Company Address List in 2022	SMTp 2023

).

2.3. Hypothesis Building

If we look at the population change from 2014 to 2023, we can see that the population of Seoul, the core of the SMA, has been declining, while the population has been concentrated in some spaces on the outskirts of the SMA. It is a known fact that from 2017 to 2021, housing prices in existing urban spaces in the SMA skyrocketed, causing a large number of citizens to move to the outskirts of the SMA, as well as to move into new towns under public development, such as Dongtan and Hanam New Town. Population growth is also noticeable in the southeastern part of the SMA, which has not been supplied with any new towns but is known to have a growing population due to sprawl since the 2010s (Figure 5).

Housing supply through new urban development is considered the first significant cause of population growth around the target area. In particular, the housing supply by public development can significantly affect relocating the population of Seoul, where housing prices are high, as it provides relatively cheap housing by developing inexpensive land in or near development-restricted areas where private development is restricted. Not only does this have a sprawl effect, but it can also intensify sprawl by increasing development pressure on the periphery of public development sites. In South Korea, it is typical for publicly led new urban development and urban redevelopment areas to be designated as District Unit Plan Zones (DUPZ), most of which have been designated since 2000. The data visualization shows that DUPZs and target areas have similar distributions of rapidly growing populations (over 1000 people) (Figure 6).

The second major cause is population spread to areas around new roads. As the metropolitan area’s total population increases, the population that cannot be accommodated by public housing development is expected to move to areas with relatively low land prices and high accessibility to the center of the urban area. Therefore, the number of new interregional roads built since 2014 is likely related to population growth. However, in Korea, highways are generally only accessible through limited toll plazas, so they should be considered separately [39].

A third possible explanation is that the supply of new cities and roads increases development pressure in non-urban areas, increasing population in areas with large amounts of unmarketed flat land, such as farmland and open space. This evidence shows that the planned supply of new towns and roads can lead to unplanned population growth. On the other hand, the development of sloping forests in South Korea is relatively strictly regulated, so it is unlikely to be related to population influx areas.

This study seeks to prove the following hypotheses to identify predatory sprawl in the SMA.

Hypothesis 1.

During 2014–2023, the population of the SMA will increase more in areas with more potential for development gains in 2014. For example, there will be more growth in areas with lower land prices, in and around planned developments, and adjacent to the greenbelt.

Hypothesis 2.

During 2014–2023, the population of SMAs will grow more in places with poor transportation in 2014 and improved transportation since then. For example, the lower the road rate and highway accessibility around 2014, the greater the area of new roads in the grid, the higher the growth.

Hypothesis 3.

During 2014–2023, the SMA population increased more in areas where non-urban land use predominates in 2014. For example, areas with a higher proportion of agricultural land and undeveloped open space would have grown more.

3. Research Method and Variables

3.1. Spatial Regression Analysis Using Geoda

Spatial regression analysis was used to test the research hypotheses. Spatial regression analysis is divided into the general spatial regression model (GSRM), which considers the spatial correlation between points in the sample, and the spatial quality model, such as GWR, which assumes that the impact of independent variables differs between points. Spatial lagged models (SLM) and spatial error models (SEM) are commonly used as general spatial regression models. There is also the spatial autoregressive combined model (SACM), which includes both spatial lag and spatial error, and the spatial Durbin model (SDM), which includes spatial lag, autocorrelation, and time difference with independent variables.

The spatial lag model, also known as spatial autoregressive regression (SAR) or Spatial autoregressive model (SAM), considers that each sample’s dependent variable is spatially correlated with the dependent variable of neighboring samples. The spatial lag is a spatially weighted matrix of the difference between each sample and its neighbors and generally reflects the value of the nearest sample, while the farther away the sample is, the less it is reflected or not reflected. The spatial error model is a model that considers the spatial correlation of errors that non-spatial econometric models cannot explain. In the structure below, W is the spatial weight matrix, which represents the relationship between points in the sample. The spatial weight matrix expresses the influence of neighboring samples through cross-standardization, and its structure can be applied in various ways, such as adjacency, recency, and economic weight. In the following equation, if $W_2$ is zero, it is a spatial lag model; if $W_1$ is zero, it is a spatial error model. If both are zero, it is a non-spatial econometric model; if neither is zero, it is a SAC.

$$\mathrm{y}=\mathsf{\rho }{W}_{1}y+X\beta +\mu$$

$$\mathsf{\mu }=\mathsf{\lambda }{W}_2\mu +ϵ$$

$$\mathsf{ϵ}~\mathrm{MNV}\left(0,{\sigma }^2{I}_n\right)$$

Specifically, the analysis program used Geoda, an open-source program provided by The Center for Spatial Data Science at The University of Chicago. The entire data were integrated into a single grid dataset, a spatial weighting matrix was created, and spatial regression analysis was performed.

3.2. Variables and Basic Statistics

The dependent variable in this study, POP_C, is the population change from 2014–2023 by grid size of 1 km × 1 km. The boundaries of all grids were aligned with POP_C, which follows the standard grid frame of Korea. For non-grid data, we recalculated the values for each grid accordingly. For example, in the case of land use, the area ratio of each land use in each grid was calculated and analyzed as a variable. The spatial weighting matrix was calculated using the distance weight method because it is based on grid data that are uniform in size and shape and placed continuously. The distance band had to be set without a clear standard, but we compared the results using 1.5 km and 2.5 km. With a bandwidth of 1.5 km, each grid only considers the relationship with its neighboring grids. With 2.5 km, it also considers the grids immediately outside it, so each grid analyzes a range of 5 km horizontally and vertically. They are sufficient to consider the impact of spatial agglomeration, as even large-scale residential projects are rarely more than 2 km on one side.

First, we constructed independent variables to test each hypothesis (

Table 2. Basic statistics.

Variable Category	Variable Name	Unit	Description	Min.	Mean	Max.	Std.
Independent Variable	POP_C	Persons	Grid Population Change from 2014 to 2023	−41,691.0	−39.5	31,129.0	2485.2
Explaining Variables	LP14	KRW 1000 (=1 USD)	Official Land Price in 2014	0.0	267.1	15,656.0	776.9
	DP_R	%	Area Ratio that Development Plan Existed in 2022	0.0	10.0	100.0	21.6
	GB_R	%	Greenbelt Area Ratio in 2022	0.0	13.3	100.0	31.0
	ORD_R	%	Area Ratio of Road Constructed by 2013	0.0	4.4	29.8	5.5
	NRD_R	%	Area Ratio of Road constructed between 2014–2022	0.0	0.5	20.9	1.6
	HWA	km	Distance to the Nearest HW entrance in 2016	0.0	9.9	61.3	12.4
	AG_R	%	Agricultural Land Area Ratio in 2010	0.0	24.6	100	26.3
	BL_R	%	Bare Land Area Ratio in 2010	0.0	3.4	100	5.7
Controlling Variable	NCOM	Unit	Counts of Companies founded after 2014 start	0.0	7.5	1156.0	35.6
Adjusting Variable	POP14	Persons	Grid Population in 2014	0.0	2389.6	53,257.0	6836.6

).

LP14, the independent variable used to test Hypothesis 1, is the average value of the grid’s public land price (per square meter) in 2014, with a unit of KRW 1000. In the original data, some develop-impossible areas, such as mountainous areas, have no value, but this study replaced them with 0 for convenience. In reality, land prices in these areas are mostly below KRW 1000, so POP_C is close to KRW 0. DP_R is the proportion of the area in each grid with land development plans, and GB_R is the proportion of the area in the greenbelt.

To test Hypothesis 2, NRD_R was extracted from the National Numerical Map based on the centerline of newly named roads since the beginning of 2014 and calculated as the ratio of road area for each grid. In the same way, we processed the roads that were named until the end of 2013 to use the ORD_R variable. In addition, HWA, highway accessibility, had a grid size of 500 m × 500 m in the original data, which is smaller than the grid size of 1 km × 1 km of the dependent variable, so the minimum value in each grid of POP_C was selected as a representative value.

AG_R is the agricultural land ratio, and BL_R is the bare land ratio used to test Hypothesis 3. Land cover data were intersected to each grid to calculate the percentage of land cover in each grid area of 1 square kilometer.

As an additional control variable to increase the model’s explanatory power, we added NCOM, the number of firms established after 2014. This control variable is necessary because, in addition to the spatial characteristics of interest in this study, the number of jobs in firms is known to impact population change significantly.

POP14, referring to the population in 2014, is used as a moderating variable. Since the dependent variable in this study is the amount of population change rather than the rate, it is essential to note that the more populated grid will have a more significant amount of population change. For the same reason, it is rare for a sparsely populated area to suddenly become heavily populated without an artificially ample housing supply, such as a new town. This relationship between POP_C and POP14 makes POP14 a poor explanatory variable, as it multiplies not only the increase in POP_C but also the decrease. It is a typical characteristic of a moderating variable. Alternatively, the rate of change rather than the amount of population change could be used as the dependent variable. However, this would be inadequate to account for the absolute increase in population and the low-density spread of people in high-density spaces, which are essential characteristics of sprawl.

The underlying statistics of the variables show extreme minimum and maximum values, especially for the dependent variable POP_C. They are the results of errors in some of the 2014 and 2023 population grids that are amplified when calculating the difference between the grids. However, these extreme cases are few. With grid data consisting of thousands of samples or more in a continuous quadratic plane, it is difficult to individually remove or correct for these outliers at the expense of explanatory power. We can also see that the population growth grid has a negative mean value even though the metropolitan area’s total population has increased. This is because the number of depopulated grids is much higher than that of populated grids. LP14 also has a much larger maximum value of 15,656.0 compared to an average value of 267.1. The variables for the area ratio are formed from 0 to 100, but the average value is small, around 10. This is because the frequency of zeros is significantly higher than that of values close to 100.

4. Result

4.1. Models without Moderating Variables: Model(1-N)

When conducting a spatial statistical analysis of the entire SMA without the moderating variable POP14, we find that all three hypotheses are significant: predatory development has been practiced throughout the SMA in the last decade, and there has been spatial sprawl in areas where it has been facilitated. To improve the explanatory power and check the robustness of the model, we run non-spatial and SLM SEM models with weight bandwidths of 1.5 km and 2.5 km. The results are broadly similar, with differences in explanatory power. The explanatory power is 0.134 for the non-spatial Model(1-0), while the spatial models range from 0.171 to 0.190. LP14 and ORD_R are consistent across all models with significantly negative coefficients, while DP_R, GB_R, NRD_R, HWA, AG_R, and BL_R are significant with positive coefficients. The control variable NCOM is also significant (

Table 3. Model results without moderating variables.

Model	Model(1-0)	Model(1-1)	Model(1-2)	Model(1-3)	Model(1-4)
Method	Non-Spatial	SLM		SEM
Weight Bandwidth	-	1.5 km	2.5 km	1.5 km	2.5 km
LP14	−0.19 ***	−0.15 **	−0.13 **	−0.26 ***	−0.22 ***
	(0.05)	(0.04)	(0.5)	(0.06)	(0.06)
DP_R	14.85 ***	10.51 ***	10.52 ***	13.28 ***	13.04 ***
	(1.35)	(1.32)	(1.34)	(1.54)	(1.52)
GB_R	2.02 *	1.59 *	1.79 *	3.00 **	3.75 **
	(0.80)	(0.78)	(0.79)	(1.12)	(1.18)
ORD_R	−128.18 ***	−101.31 ***	−104.15 ***	−115.48 ***	−114.59 ***
	(6.42)	(6.30)	(6.38)	(7.38)	(7.36)
NRD_R	394.67 ***	330.05 ***	348.99 ***	369.24 ***	379.65 ***
	(16.36)	(16.09)	(16.24)	(17.86)	(17.47)
HWA	0.278	0.18	0.39	1.76	2.87
	(2.14)	(2.07)	(2.10)	(3.15)	(3.53)
AG_R	5.27 ***	3.89 ***	3.90 ***	6.07 ***	6.45 ***
	(0.95)	(0.92)	(0.94)	(1.20)	(1.21)
BL_R	19.69 ***	15.27 ***	15.90 ***	21.27 ***	22.38 ***
	(4.40)	(4.27)	(4.32)	(4.99)	(4.94)
NCOM	5.25 ***	5.25 ***	5.30 ***	6.33 ***	5.81 ***
	(0.83)	(0.80)	(0.81)	(0.87)	(0.84)
Constant	−61.66	−43.13	−50.87	−133.31	−182.28 *
	(54.01)	(52.28)	(52.94)	(73.33)	(78.51)
Spatial W	-	0.35 ***	0.40 ***	-	-
		(0.02)	(0.02)
LAMBDA	-	-	-	0.37 ***	0.44 ***
				(0.02)	(0.02)
N	9808	9808	9808	9808	9808
R2	0.134	0.189	0.171	0.190	0.173

( ): standard error, ***: p < 0.001, **: p < 0.01, *: p < 0.05.

).

Regarding Hypothesis 1, we can see that there was more population growth in places with lower land prices (LP14), in or adjacent to planned development (DP_R), and with a higher percentage of greenbelt area (GB_R). Typically, we expect these neighborhoods to have higher land prices as development occurs. However, this is open to interpretation, as places with low land prices or high greenbelt area ratios may reflect low development potential. Planned development has a much larger coefficient than the other variables. The provision of a large number of housing units by planned development has a strong effect on demographic change.

Regarding Hypothesis 2, we can see that the population grew where there were fewer existing roads (ORD_R) and more new roads (NRD_R). However, highway accessibility does not show a significant relationship. In other words, areas with a high concentration of existing roads lose population. This is an effect of the surge in house prices in the metropolitan area over the five years since 2017, which pushed people out of the metropolitan area. The new roads would have made it easier for these people to move, but they would have also increased their commuting distances. The effect of new roads is three times that of existing roads, which is significant because new roads are rarely opened.

Regarding Hypothesis 3, we can see that more population growth occurred where the agricultural land ratio (AG_R) and the bare land ratio (BL_R) were high. The coefficient of the bare land ratio is three to four times larger than that of the agricultural land ratio, indicating that it is relatively easier to attract the population through the development of bare land because there are more significant regulatory barriers to agricultural land development. However, there is a small effect in agricultural areas, suggesting we should be wary of sprawl.

4.2. Models with Moderating Variable: Model(2-N)

The model’s explanatory power is further increased by adding the moderating variable POP14 to the main explanatory variables. The non-spatial model and SLM model SEM model were established as the same as the non-moderating variable model but with two weight bandwidths of 1.5 km and 2.5 km. The explanatory power is 0.311 for the non-spatial model, Model(2-0), and the spatial models range from 0.325 to 0.372, which is relatively high for a regression model using large-scale population grid data. Some of the explanatory variables change the direction of the coefficients from the results of the no-control model to the opposite direction or become insignificant. Consistently across all models, all explanatory variables are either significant or insignificant with positive coefficients, and all significant moderators are significant with negative coefficients (

Table 4. Model results with moderating variables.

Model	Model(2-0)	Model(2-1)	Model(2-2)	Model(2-3)	Model(2-4)
Method	Non-Spatial	SLM		SEM
Weight Bandwidth	-	1500 m	2500 m	1500 m	2500 m
LP14	0.91 ***	0.92 ***	0.93 ***	1.16 ***	1.07 ***
	(0.06)	(0.06)	(0.06)	(0.07)	(0.07)
DP_R	27.88 ***	23.87 ***	24.49 ***	29.31 ***	29.12 ***
	(1.44)	(1.44)	(1.46)	(1.62)	(1.58)
GB_R	1.76 *	1.30	1.53 *	1.21	1.46
	(0.70)	(0.72)	(0.73)	(1.08)	(1.09)
ORD_R	−11.04	2.79	0.74	−4.49	−4.04
	(6.40)	(6.30)	(6.37)	(7.21)	(7.16)
NRD_R	345.96 ***	300.43 ***	315.78 ***	326.91 ***	336.18 ***
	(16.29)	(16.15)	(16.28)	(17.66)	(17.25)
HWA	5.19 **	4.89 **	5.13 **	6.65 *	7.71 *
	(1.92)	(1.88)	(1.90)	(3.09)	(3.29)
AG_R	−0.63	−0.89	−0.91	0.09	0.31
	(0.88)	(0.87)	(0.88)	(1.14)	(1.13)
BL_R	13.99 **	8.99 *	10.17 *	14.77 **	17.50 **
	(4.25)	(4.18)	(4.22)	(4.79)	(4.70)
LP14*POP14	−0.00004 ***	−0.00004 ***	−0.00004 ***	−0.00005 ***	−0.00005 ***
	(0.000003)	(0.000002)	(0.000002)	(0.000003)	(0.000003)
DP_R*POP14	−0.002 ***	−0.002 ***	−0.002 ***	−0.002 ***	−0.002 ***
	(0.0001)	(0.0001)	(0.0001)	(0.0001)	(0.0001)
GB_R*POP14	−0.003 ***	−0.003 ***	−0.003 ***	−0.003 ***	−0.003 ***
	(0.0003)	(0.0003)	(0.0003)	(0.0003)	(0.0003)
NRD_R*POP14	−0.02 ***	−0.02 ***	−0.02 ***	−0.02 ***	−0.02 ***
	(0.002)	(0.002)	(0.002)	(0.002)	(0.002)
HWA*POP14	−0.01 ***	−0.01 ***	−0.01 ***	−0.01 ***	−0.01 ***
	(0.002)	(0.002)	(0.002)	(0.002)	(0.002)
AG_R*POP14	0.0002	−0.0004	−0.0003	−0.0005	−0.0005
	(0.0002)	(0.0003)	(0.0003)	(0.0003)	(0.0003)
BL_R*POP14	−0.004 ***	−0.004 ***	−0.004 ***	−0.004 ***	−0.004 ***
	(0.0006)	(0.0006)	(0.0006)	(0.0006)	(0.0006)
NCOM	−2.53 **	−2.09 **	−2.17 **	−1.72 *	−1.72 *
	(0.76)	(0.75)	(0.75)	(0.79)	(0.79)
Constant	−241.98 ***	−225.03 ***	−234.08 ***	−300.24 ***	−339.26 ***
	(48.58)	(47.61)	(48.11)	(70.15)	(72.53)
Spatial W	-	0.27 ***	0.28 ***	-
		(0.01)	(0.02)
LAMBDA	-	-	-	0.43 ***	0.47 ***
				(0.02)	(0.02)
N	9808	9808	9808	9808	9808
R2	0.311	0.338	0.325	0.372	0.346

( ): standard error, ***: p < 0.001, **: p < 0.01, *: p < 0.05.

).

As with the unadjusted models, the adjusted models confirm that there has been sprawl in the SMA over the past decade, with an increased population in spaces prone to predatory development. However, including the control variable POP14 allows for a more detailed interpretation.

Regarding Hypothesis 1, the higher the land price (LP14), the greater the population growth. However, the negative coefficient of the intercept of the moderating variable (POP14), LP14*POP14, indicates that this phenomenon is dominant in less populated areas. In other words, more populated areas may lose population to higher land prices, while less populated areas gain more population to higher land prices within the region. This result is more realistic than Models(1-N) because land prices somewhat reflect future development potential. This result is more realistic than Model(1-N) because land prices proxy for future development potential. It can be interpreted as a population outflow from a grid to a relatively high land price area within a sparsely populated area rather than simply moving to a lower area.

In the case of DP_R and GB_R, the results are the same as those of Models(1-N) in sparsely populated areas. However, the relationship may disappear or change in direction in more populated areas. In other words, the effect of development plan areas or greenbelts in sparsely populated areas increases with population but decreases as population increases. However, DP_R only turns positive to negative once the population exceeds 15,000 people, while GB_R turns negative when the population exceeds about 500 people.

Regarding Hypothesis 2, the coefficient of existing roads (ORD_R) is no longer significant, as found in Models(1-N), while the coefficient of new roads (NRD_R) is still significant. Instead, highway accessibility is significant, meaning that variables related to population outflows are more significant than inflows in the moderating variable model. Also, although the moderator NRD_R*POP14 is negative, its coefficient is much smaller than that of NRD_R, suggesting that the effect of NRD_R may become smaller as the population increases. However, the positive relationship does not reverse to a negative relationship. In other words, the more sparsely populated the area, the greater the in-migration effect of new roads. It is consistent with common sense.

Regarding Hypothesis 3, the agricultural land ratio (AG_R) is insignificant, and the bare land ratio (BL_R) is still significant. Considering the moderating variables, AG_R*POP14 is not significant, and BL_R*POP14 is significant. It implies that the moderating variable model does not confirm the in-migration effect of agricultural land. At the same time, bare land has an in-migration effect and a larger in-migration effect in less populated areas.

4.3. Heterogeneity Analysis by Sub-Regions

A subset of subregions was analyzed to check the spatial heterogeneity of the analysis results. Namyangju City, Yangpyeong County, Gwangju City, and Yongin City in Gyeonggi-do, known to have intensified sprawl within the SMA, were selected. As a result, the models were similar to the results for the entire metropolitan area. In the model for the SMA, the model with the SEM and bandwidth 1500 m options had the highest explanatory power with no difference in results, so only the SEM and 1500 m models are presented for the sub-region analysis. Also, moderating variables were not applied for the convenience of comparison and interpretation. In the case of Yongin City, the GB_R variable was excluded because there is no greenbelt in the city (

Table 5. Model results for heterogeneity test.

Model	Model(3-1)	Model(3-2)	Model(3-3)	Model(3-4)
Sub-Region	Namyangju City	Yangpyeong County	Kwangju City	Youngin City
Population in 2014	636,256	105,379	298,858	961,026
Population in 2022	737,353	122,323	391,462	1,074,971
Population Increase (Increasing Rate)	101,097	16,944	92,604	113,945
	(15.9%)	(16.1%)	(31.0%)	(11.9%)
LP14	−1.50 **	−0.43 **	−0.65	−0.31
	(0.49)	(0.14)	(0.42)	(0.25)
DP_R	20.52 ***	0.18	34.39 ***	8.20 **
	(4.16)	(0.34)	(4.30)	(2.89)
GB_R	−1.90	−0.46 *	−1.14	-
	(1.64)	(0.20)	(1.05)
ORD_R	−172.28 ***	12.74 ***	−45.90	−60.72 **
	(30.88)	(3.42)	(23.84)	(18.93)
NRD_R	638.34 ***	32.00 ***	287.23 ***	590.62 ***
	(40.10)	(8.02)	(49.96)	(52.49)
HWA	2.10	−0.60	−1.68	4.79
	(16.29)	(0.59)	(12.91)	(10.08)
AG_R	11.13 *	−0.16	4.05	−2.12
	(5.53)	(0.25)	(3.24)	(2.06)
BL_R	21.39	2.66	−31.03 **	63.37 ***
	(13.7)	(2.50)	(11.25)	(12.58)
NCOM	−8.57	−21.79 ***	4.42 **	26.24 ***
	(11.49)	(4.52)	(9.46)	(4.16)
Constant	87.29	10.43	96.51	−1.79
	(167.88)	(8.63)	(98.52)	(92.96)
LAMBDA	0.08	0.24 ***	0.25 ***	−0.25
	(0.08)	(0.05)	(0.07)	(0.08)
N	526	996	509	646
R2	0.443	0.110	0.210	0.271

( ): standard error, ***: p < 0.001, **: p < 0.01, *: p < 0.05.

).

Namyangju, Yangpyeong, and Gwangju are located on the eastern side of the SMA and have been known as representative areas of suburbanization in the SMA. These areas are separated from Seoul by a greenbelt, and residents commute to Seoul via highways, arterial roads, and urban railways. These areas have some of the lowest land prices in the SMA and have been served by new trunk roads over the past decade. Namyangju and Yangpyeong counties have each seen population growth of about 16%, while Gwangju City has seen a 31% increase.

Yongin is located in the southern part of the SMA and is a case study area that illustrates the consequences of weakening development regulations around new towns. Rapid sprawl has occurred in the undeveloped areas surrounding the existing large new towns of Bundang, Dongtan, and Gwanggyo. The population grew by 113,945 people, or 11.9% in percentage terms.

The analysis of these subregions shows that the effect of NRD_R is significant in the same direction as the SMA results in all four subregions, and LP14, DP_R, and ORD_R are also significant in the same direction as the SMA full model in at least two regions. It suggests that new roads can strongly incentivize sprawl in these subdivisions, further supporting Hypothesis 2. LP14 and DP_R also support Hypothesis 1, but GB_R is insignificant or has the opposite result. Regarding Hypothesis 3, AG_R and BL_R are only partially significant or have opposite results, so it is challenging to conclude Hypothesis 3 in subregions.

5. Discussion and Conclusions

This study examines the population change in SMAs in Korea over the last ten years and tries to determine whether the sprawl phenomenon has the aspect of predatory development. For this purpose, a spatial regression analysis was conducted using the population grid of the SMA and independent variables that can explain the value of the population grid. In particular, we tried to establish a model that considers the spatial heterogeneity of SMAs by inserting control variables. As a result, we found that the population growth from 2014 to 2022 is concentrated in spaces that are not subject to planned development but are expected to have enormous development benefits and are easy to develop.

Hypothesis 1, that places with tremendous development profit potential are more likely to be developed predatorily, is statistically significant for three factors: land price, development plan, and greenbelt, but there are differences in the details. First, while it is true that lower land prices attracted people, it can be interpreted as a result of more out-migration from existing population centers with very high land prices. They moved to areas with lower land prices overall, but within those areas, they settled in areas with relatively higher land prices. The evidence is that the coefficient of LP14 in Model(2-N), which includes the moderating effect of POP14, turns positive. These results suggest that land prices are a proxy for development potential. Therefore, spaces that can attract the population within undeveloped areas already have relatively high land prices within those areas.

Our finding that a development plan (DP_R) is a strong determinant of in-migration supports Hypothesis 1, too. The granting of a formal development plan formalizes the realization of development potential, and the surrounding area may develop further in anticipation of future development. It can lead to population influx. In the case of Yangpyeong County and Gwangju City, which did not have large-scale new town projects, DP_R was not statistically significant in Yangpyeong County but was statistically significant in Gwangju City.

Hypothesis 2, which states that improving transportation infrastructure in transportation-inconvenient areas (NRD_R) induces population inflow, has been addressed in previous studies. However, this study has great significance in intuitively confirming it through grid-based spatial analysis. In particular, the accessibility of highway toll gates was not significant in Model(1-N). However, in Model(2-N), which considered the moderating effect of population, there was a significant result: the farther away the toll gate is, the greater the population inflow. When the coefficients of ORD_R, NRD_R, and HWA are combined, it can be seen that sprawl spreads from places with transportation infrastructure to places where transportation is inconvenient and then to places where transportation is improved.

For Hypothesis 3, which states that population inflows to areas with predominantly non-urban land use will be greater, we found a significant effect of bare land (BL_R) and agricultural land (AG_R). However, the effect of agricultural land was less potent than that of bare land. In the research process, we also analyzed forests and grasslands, which were insignificant. It may be because Korea still has strong development regulations on farmland and mountainous areas. However, since the results of this study confirm the spatial correlation of population inflows, it is reasonable to expect that the development potential of agricultural land and forests in areas with high concentrations of bare land will increase.

This study differs significantly from previous studies because it uses in-depth grid data. As a result of using grid data, first of all, we found that the explanatory power is not so high. It is an advantage of grid data that the independent variables are composed of spatial and intuitive variables. However, the distribution of the dependent variable is also greatly diversified, which is thought to reduce the explanatory power. For example, if we had analyzed the data at the administrative district level as in previous studies, the number of samples would have been reduced. However, their distribution would have been spread over a larger area, and most of the effect would have been explained by spatial agglomeration. However, in practice, there will be overpopulated and underpopulated areas within each administrative region, making it more challenging to draw intuitive conclusions.

In addition, it is significant that the study examined the moderating effect of POP14. The main reason for the decline in explanatory power was the large area of the entire SMA and the difficulty of presenting a single model with a large local population distribution. In particular, if the population of the grid is large, both the increase and decrease in the population are bound to be large, so it was appropriate to use population as a control effect. As a result of the analysis, the explanatory power increased by about 20%, which can be considered sufficient explanatory power to trust the effect of other variables, and it is expected that dynamic implications can be obtained by comparing Model(1-N) and Model(2-N).

The results of this study identify which areas have the highest development potential and, therefore, attract more people. It has important implications for future urban planning policies. On a broader scale, areas with low land prices and improving transportation infrastructure should be prepared for population influx and subsequent sprawl. Locally, if there are areas where land prices have risen sharply because development potential is reflected in advance, it is necessary to check whether urban planning is adequately prepared. Especially in areas where development plans have been granted, it is necessary to monitor not only the area but also the surrounding areas. If there is a high percentage of bare land, even more attention is needed.

SMA is experiencing the in-ward sprawl. As Seoul grew in the 1960s, she made urban planning of the urban–region system preserving overall green areas. Seoul successfully preserved the green area while accommodating more than 10 million people in dozens of urban cores. In recent years, with the strong deregulation wave, architectural developers have been making a big fortune by developing green areas between the urban cores. Since the green areas between the urban cores are very cheap because of the regulation, developers can make a big fortune by converting the un-developable land into developable land with manipulation or comptonization of the planning regulation. Consequently, once well-planned cities are getting worse and becoming similar to the sprawl cities with congestion and pollution.

This is a kind of paradox of urban planning. Urban planning designates some areas for development and other areas for preservation. Most preservation areas are environmentally sensitive areas or preservation-worth areas (i.e., ecological value, clean water protection, etc.). However, developers are eager to develop such preserved areas mainly because the land is most profitable. That is, the preserved land is one of the cheapest areas because of urban planning. But, if the land becomes developable, the value increases tremendously.

This paper tried to model the spatial pattern of the recent in-between sprawl occurring in well-planned cities with the deregulation trend. The result of this paper helps to locate the most vulnerable places from predatory developers and provides a planning implication for maintaining sustainability.