Forecasting the urbanization dynamics in the Seoul metropolitan area using a long short-term memory

Abstract

Over the past half century, the Seoul metropolitan area (SMA) has experienced rapid urbanization. Urban development and population growth within the SMA have caused various problems, such as a lack of affordable housing, traffic congestion, and socioeconomic inequality between the SMA and the rest of the country. As a solution, growth control was adopted, but it resulted in increasing housing prices within Seoul. In late 2018, skyrocketing housing prices forced Seoul’s government to abandon its growth-control policy and announce large-scale “new-town” projects planned outside of the city’s urban growth boundary. The primary purpose of this research is to predict future urbanization dynamics by utilizing the long short-term memory (LSTM)–based prediction model. The secondary purpose is to identify the influential driving factors in urbanization that can help policy makers develop evidence-based, informed strategies. To predict future urbanization’s spatial patterns in the SMA, LSTM models have been estimated under two scenarios: (A) assuming that current urbanization trends and contributing factors will remain consistent in the future and (B) considering new development plans’ impacts. A comparison of the modeling results indicates that the government-driven new-town projects will help urbanize 55.8% more land by 2030. The variable influence analysis also reveals that strong growth-control measures may be necessary for areas with higher employment and homeownership rates to control rapid urbanization. However, housing supply and economic growth–related policies in Seoul’s suburbs would help attract the city’s population to the outskirts. The LSTM-based model yields an accurate and reliable spatial prediction in the form of visual maps, and its graphic results will assist policy makers greatly in developing effective strategies for smart urban growth management.

Keywords

Urbanization urban land-use change model long short-term memory sprawl Seoul

Introduction

Over the past half century, South Korea has experienced rapid urbanization. Its current urban population is 84.5% of its national population, more than double what it was in 1968 (40.7%) and quadruple that of 1950 (21.4%) (Kim et al., 2003) (Supplementary Figure 1). Today, nearly half of South Koreans live within the Seoul metropolitan area (SMA), and the population concentration continues, resulting in significant socioeconomic inequality between the SMA and the rest of the country. The City of Seoul, which is much smaller than the SMA, has suffered from overcrowding and a lack of developable land. In 2010, over 12% of Seoulites lived in substandard housing conditions, and the homeownership rate also was low (44.6%) compared with the national average of 55.6% (Ha, 2010).

To reduce overcrowding within Seoul, the government built master-planned satellite cities. Government-driven, large-scale, master-planned satellite cities called “new-towns” in Korea were implemented in two phases, with the first in the 1990s and the second in the 2000s. The supply policy helped improve housing affordability within the SMA, but it simultaneously was the catalyst for accelerating population concentration in the SMA and for increasing socioeconomic inequality between the SMA and the rest of the country.

After being inaugurated in 2017 on a liberal political platform, President Moon moved in the opposite direction by supporting small-scale urban regeneration and limiting large-scale housing development within the SMA. The third phase of the new-town plan includes the following (Yoon, 2019): (1) 150,000 new houses through development of 34.3 km² in the SMA and (2) construction of 31 stations for three new rail transit lines, called the Great Train Express (GTX). To avoid commuting nightmares, the new-towns during the third phase are set to be located closer to Seoul than those from the previous two phases. This can be achieved only by redesignating a significant amount of land currently designated as greenbelt land as “developable” land. Environmental concerns significantly eroded public support for the new-town plan’s third phase, but the government does not seem inclined toward changing its direction so far.

As the third phase is expected to influence future urban growth policy significantly, it is important to forecast its impact on urbanization. In this study, the spatial transitions for future urbanization are predicted under two scenarios: (1) Scenario A assumes that current urbanization trends and contributing factors will remain consistent in the future, and (2) Scenario B considers impacts from the recently announced development plans, including the six new-town projects and construction of the 31 GTX stations. The urbanization patterns projected under these two scenarios will be compared with draw implications for future urban growth policies. Although the new-town projects are not under construction yet, a new-town is master-planned under government supervision with predetermined population size and land area, which are not changeable after the public announcement. Thus, this research utilizes the exact location and size of the future new-town areas.

This study has two goals: (1) to predict future urbanization dynamics in the SMA by utilizing the long short-term memory (LSTM)–based prediction model and (2) to identify the influence from each urbanization variable that can help policy makers develop evidence-based, informed planning strategies. First, to simulate urban land-use transitions accurately and quantify the effect from each variable on future urbanization, this research utilizes a neural network model, that is, the LSTM-based model. Using time-series data from 1990 to 2010, we simulated future urbanization dynamics, analyzed prediction output accuracy, and compared urbanization trends projected for the SMA under two different scenarios. The second research goal is to quantify the effect from each driving factor contributing to urbanization. The outcomes from this research are expected to provide logical foundations for effective policy decisions that eventually will contribute to evidence-based decision-making and sustainable urban growth management.

Please find the full version of introduction and literature review with the background description in the Supplementary Materials.

Literature review

Urbanization trends in Korea

The urbanization pace in developing countries is much faster than in developed countries, and the urbanization trend in Korea is no exception. It took 90 years for the United States to reach an urbanization rate of 80% from 40%, while it took Korea only 20 years to reach 80% from 40 (Henderson, 2002). In 2010, the urbanization rate in Korea exceeded 85%, which is significantly above the Organization for Economic Co-operation and Development’s (OECD) average of 47.1%. Although the Korean government has endlessly attempted to stabilize housing prices in various ways, the housing demand is still not satisfied. Approximately half of the total population in South Korea is living in the SMA; this high concentration makes it difficult to supply affordable housing near major employment centers. To stabilize skyrocketing housing prices within Seoul, the government has tried all solutions available, including price-control and land regulations, gross real estate tax, and the development of new satellite cities (Kim and Han, 2012; Kim and Kim, 2000). The government effort, however, has fallen short most of the time.

With rapid population growth concentrated in Seoul and nearby areas, the government established its first Seoul Metropolitan Area Readjustment Planning Act in 1982. Unfortunately, various government efforts to mitigate the population concentration have been mostly inconsistent, failing to disperse the population into the rest of the country. The recent policy U-turn by Korea’s liberal party toward neutralizing the greenbelt for new satellite cities was another setback for environmentalists.

LSTM-based spatial pattern prediction model

The causes and consequences of urbanization and its impact on land-use transformation have been studied. The most widely used method to predict future land-use patterns is a combination of regression analyses and geographic information systems (GIS). Regression-based models are powerful in identifying functional relationships between variables. GIS-based models can deliver their outcomes in a form of spatially explicit and easy-to-understand digital maps (Turner and Meyer, 1991; Verburg et al., 2006). Combining these advantages led to the development of elaborate simulation packages, such as UrbanSim (Waddell, 2002) and Cellular Automata models (Batty et al., 1999; Liang et al., 2018). Policy makers and planners who are not familiar with statistics and economic theories can understand spatial analysis results much more easily with graphical outputs. More recently, neural network–based models, such as the land transformation model, have risen in popularity as alternatives for predicting land-use dynamics (Grekousis et al., 2013; Liu et al., 2021; Newman et al., 2016). Pijanowski et al. (2002) proved that artificial neural network (ANN)–based models are useful for handling nonlinear patterns in data and generate a more than 40% accuracy rate, which is higher than traditional regression models. Furthermore, by improving computer performance, many neural network–based models have been developed to increase model accuracy (Pontius et al., 2008; Tayyebi et al., 2013).

As proven by many previous studies, LSTM has powerful predictive ability compared to traditional statistical models. Nghiep and Al (2001) developed an ANN-based prediction model to forecast sale prices of private properties in Rutherford County in Tennessee and compared the accuracy with multiple regression analyses. Furthermore, Azari et al. (2019) proved that LSTM-based prediction model performed about 5% better than ARIMA to predict user traffic in a cellular network. However, most of them were focused on traffic flows and real estate price prediction using time-series analyses (Liu et al., 2021). Newman et al. (2016) and Lee and Newman (2017) developed an ANN-based model to predict urban spatial pattern dynamics (vacancy) and verified the model’s performance with four different methodologies. The prediction accuracy of the ANN-based model was only 54.7%. Furthermore, many prediction models were performed under the assumption that the current trends will remain consistent in the future. Thus, this research utilized more accurate LSTM-based spatial prediction model to predict future urbanization dynamics and evaluate the impact of new-town developments on future urbanization in the SMA.

Literature gaps and research objectives

In diverse fields––including economics, statistics, planning, and geography––significant research efforts have been made to predict future spatial pattern transition scenarios and examine the causes and consequences of spatial transformation. Accurate projection of future urbanization provides invaluable information that helps policy makers develop the best planning strategies. The most commonly used spatiotemporal prediction models have been GIS-based regression models, which frequently are limited by a lack of accuracy in the assessment process and, thus, relatively low reliability. Even ANN-based prediction models, which have been accepted with 40% reliability, have proven to be more reliable than traditional models, such as the Land Transformation Model (LTM). Furthermore, while many GIS-based prediction models focus on predicting urban growth patterns, relatively little research has sought to quantify urbanization determinants’ influence. While some studies have dealt with urbanization’s causes and consequences based on statistical analyses (Liddle and Messinis, 2015; Shahbaz and Lean, 2012; Zhang et al., 2014), no such study has yet to choose LSTM-based models to analyze urbanization in the SMA, although LSTM can be used to predict land-use patterns and identify functional relationships between patterns and drivers. In this context, this study’s primary purpose is to predict future urbanization dynamics by utilizing the LSTM-based prediction model. The secondary purpose is to identify the influential driving factors in urbanization that can help policy makers develop evidence-based, informed planning strategies.

Data and methods

Study area

The study area selected for this research is the SMA, which comprises Seoul and 31 cities in the Gyeonggi province, including Incheon (Supplementary Figure 2a). This study utilizes land-use inventory data provided by the National Environmental Information Network System (NEINS) in 10-year intervals from 1990 to 2010. Based on this land-use data, built-up areas were selected and rasterized at a resolution of 500 × 500 m. Seoul, South Korea’s capital, has been the center of the Korean economy, politics, and culture for over 600 years, and the population was distributed evenly until the 1950s, when only 2.45 million (9.6%) people lived in Seoul. However, rapid economic development in the 1960’s accelerated an increase in the urban population. Mass migration from rural areas took place, with most of these migrants settling in Seoul and its suburbs. The Asia Society’s 2013 report revealed that about 13.6% of the rural population migrated to the cities—61% to Seoul—a trend that made Seoul a migrant city in the 1970’s. Seoul’s population increased continuously and reached 10.58 million (21.7%) by 2010, rising by more than 331% over the prior four decades.

After the 2010 peak, the growing population spilled over into the surrounding SMA. As shown in Supplementary Figure 2b, which presents population growth patterns from 1960 to 2018, Seoul’s population dropped steadily starting in 1991 and was approximately 9.7 million in 2018. However, at the regional level in the SMA, there was no sign of depopulation and/or decentralization. The population spilling over into new satellite cities has accelerated urbanization in the SMA and has exacerbated socioeconomic inequality between the SMA and the rest of the country. According to the Asia Society’s 2013 report, eight out of 10 rural-to-urban migrants moved to the SMA, and nearly half the national population (49.6%) is living within this area, which comprises only 11.8% of Korea’s total geographical land area (Shahbaz and Lean, 2012).

Figure 1 indicates the land pattern transitions from the SMA’s actual urbanization between 1990 and 2010 in 10-year increments. As the population increased, the total built-up area increased from 4.6% to 13.0% over 20 years. The rate of the developed area more than doubled across the Incheon and Gyeonggi regions, while the urbanized area in Seoul increased by only 20.8% during the same period. Considering that over 52% of Seoul already was developed in the 1990s, a shortage of developable land in the city led to new development on the periphery, specifically in the Incheon and Gyeonggi provinces.

Figure 1.

Historical pattern and ratio of built-up area in the SMA between 1990 and 2010 at 10-year intervals.

To solve the problems caused by heavy population concentration in the SMA, the Korean government proposed various solutions. The first balanced national development plan was established in 1984, and the third Seoul Metropolitan Area Readjustment Planning Act was announced in 2006 to decentralize the population around the capital area by 2020. Many local governments aimed to attract businesses to their cities, giving them names such as “Innovation City” and “Enterprise City.” Despite these efforts, the SMA’s overpopulation problem became acute due to a lack of public finance capacity to tackle such problems. A Ministry of Land Infrastructure and Transport survey in 2019 revealed that seven of the nine goals from the third Seoul Metropolitan Area Readjustment Planning Act had not been met (Supplementary Table 1). In the case of decentralizing the population around the capital area, contrary to the government’s expectation that the proportion of the population would decrease from 47.9% to 45.5% by 2020, it is now expected to increase to 49.7% in 2020. Under these circumstances, it is critical to develop a precise spatial prediction mechanism to facilitate more proactive decisions regarding population concentration control.

Variable selection

Considering that factors’ sources, types, and features can influence spatial prediction outputs and the model’s accuracy and performance, it is critical to identify the principal causal mechanisms contributing to urbanization. Based on the literature, this study sorted the primary causes into three categories: accessibility; economic activity; and individual socioeconomic status. These factors can contribute to rapid land development by stimulating residential, commercial, and business activities. According to the existing literature, the factors related to accessibility are associated highly with the expansion of urbanized areas. Considering that a high level of accessibility and connectivity to amenities and services attracts new residents and increases development, sites closer to highway entrances and major transportation depots are more likely to be urbanized (Antrop, 2004; Arvin et al., 2015; Atack et al., 2010; Kotavaara et al., 2011; Wang et al., 2016). Considering that proximity to highways and railway stations represents connectivity and accessibility, sites located near existing transportation lines would increase urbanized areas. The second factor associated with urbanization is economic activity level. Economic activity is associated with the ability to draw people to an area. As people follow jobs, and jobs draw people, higher employment and better employee resources can accelerate urbanization (Fu and Hong, 2011; Grant, 2012; Henderson, 2002; Jones, 1991; Sato and Zenou, 2015). Furthermore, considering that many urban areas experiencing rapid urbanization contain a large share of service industries in Seoul, sites with a larger non-secondary industry would be more likely to get developed. Finally, socioeconomic- and housing-related factors, such as population density and homeownership (housing supply), can contribute to urbanization, and these factors also are associated highly with job availability (Atack et al., 2010; Andersson et al., 2009; Buhaug and Urdal, 2013; Henderson, 2002; Kumar and Kober, 2012; Vatuk and Coleman, 1972). As people move to urban areas, housing supply, population density, and individual socioeconomic status can be affected. It is assumed that higher population density, homeownership, and education level can accelerate urbanization. With this in mind, we culled nine input factors from the literature that potentially could contribute to urbanization.

Table 1 describes each input factor, provides a description, and lists the relevant research and data source in which a causal driver in urbanization trends was utilized. Transportation data (proximity to railway stations and highways) were collected from the Korea Transport Institute. Transportation access is what differentiates Scenario A from Scenario B. With the latter, the impacts from the recently announced development plan were considered. The third phase of the new-town project includes construction of six satellite cities and 31 GTX stations for three rapid transit lines. To measure accessibility, the distance from highway and depot centers was classified into five categories using a 1 km buffer. Scenario A assumes no additional government-driven measure for housing development.

Table 1.

Driving factors in urbanization and data sources.

Domain	Input factors	Definition	Data source
Accessibility	KTX	Proximity to KTX railway stations	Korea Transport Institute
	GTX	Proximity to GTX railway stations * only for scenario B
	Highway	Proximity to highways
Economic activity	Secondary industry	Secondary industry rate (manufacturing and construction)	Korean Statistical Information Service
	Working population	Economically active population rate (age range: 15–64)	MicroData Integrated Service
	Employment	Ratio of people employed in the working-age population	MicroData Integrated Service
Housing/individual	Population density	Number of people per square kilometer	Korean Statistical Information Service
	Homeownership	Owner-occupied housing rate
	High school graduate	Percentage of people over age 25 with more than a high school diploma

Industry- and socioeconomic-related data, such as employment rate and homeownership, were provided by the Korean Statistical Information Service at the 424 census district level. Urbanization trend inventories and factor data then were rasterized at a resolution of 500 × 500 m. While several studies have examined the critical causal mechanisms that contribute to urbanization, a lack of quantification of the influence on the increase in built-up areas remains. Thus, identifying each variable’s influence is one of this study’s research goals.

Methods

LSTM modeling has four sequential steps: (1) data filtering and grid integration—the call of the input patterns (three different time frames), driving factors, and exclusive layer data are stored and rasterized using a fixed cell size (500 × 500 m) within GIS; (2) model validation—the land-use dynamics of the actual input pattern from the most recent year (2010) and the transition expected by LSTM (2010) are compared; (3) temporal scaling of the forecasted output—the amount and location of land predicted to change in the future are determined based on historical urbanization patterns and population changes; and (4) an influence assessment—the influence of each input factor on model performance is assessed to determine which variable exerts a stronger effect on urbanization.

Among many deep-learning–based prediction models, LSTM resolves issues with long-term dependencies on RNN and sufficiently reflects the correlation of historical input data. When time-series data’s time intervals are longer, a vanishing gradient problem can occur in traditional RNN, reducing learning ability (Hochreiter, 1998; Squartini et al., 2003). To overcome this issue, LSTM was developed. Although RNN and LSTM have the same structure in the form of a chain, a structural difference exists in the internal recurrent module. As shown in Supplementary Figure 3, LSTM has three gates in the internal module, enhancing the memory to remember past events more efficiently than RNN.

As shown in Figure 2, two different types of input layers were utilized to predict future urbanization trends in 2020 and 2030; built-up area data were extracted from (1) rasterized historical land-use inventories for three different time frames (referred to as input patterns) and (2) rasterized causal variables linked to a spatial location in 1990, 2000, and 2010 (referred to as input factors). Furthermore, undevelopable water bodies and wetlands were excluded from the analysis (referred to as exclusionary layers).

Figure 2.

Conceptual flow diagram showing input patterns, input factors, and output layers.

The LSTM-based model then was run to predict future urbanization trends under two different scenarios. For Scenario B, new development plans, including the six new-town developments and the construction of the 31 GTX stations, were viewed as the additional input pattern (Figure 3). Considering that the new-town projects provide a population plan as well, we also considered the population growth near the developments for Scenario B. Using input layers and exclusionary layers, we produced a map of the expected 2010 built-up area. Then two different accuracy-assessment processes, the cell-matching rate and kappa statistics, were performed by comparing the actual urbanization transition with the predicted transitions. The outcomes for all comparisons were validated, and the 2020 and 2030 urbanization patterns were predicted. Finally, each input factor’s influence on model performance was quantified.

Figure 3.

31 GTX stations and six planned new-town development areas.

Results

Possible urbanization scenarios and output statistics

Future urbanization’s spatial patterns were projected under two scenarios—one with the new development plan and one without. To predict the areas that will be urbanized in the future, we assessed the spatial input layers between 1990 and 2010. The built-up area data were pixelated by a 500 × 500 m grid as the analysis unit by dividing all the urban cells into a total of 46,850. As shown in Table 2, the number of urban cells (500 × 500 m) that transitioned to built-up areas during these two decades was +3350 (+121.8%). Under Scenario B, the number of additional urban cells (500 × 500 m) for the proposed new-towns was 136 pixels (+0.29%). Table 2 also shows the output statistics for all models and scenarios, along with the number of cells, the urbanized area’s rate, and the greatest accuracy of each training cycle. The comparison of the actual and predicted outputs provided four different values: (1) TP—actual and predicted; (2) FP—not actual, but predicted; (3) FN—actual, but not predicted; and (4) TN—not actual and not predicted. Based on the number of pixels for each value, we utilized five measures to assess prediction accuracy––the kappa coefficient, cell-matching rate (accuracy), precision, sensitivity, and specificity––as follows

Cell - matching rate = \frac{T P + F N}{T P + F P + F N + F P}, Precision = \frac{T P}{T P + F P}, Sensitivity = \frac{T P}{T P + F N}, Specificity = \frac{T N}{F P + T N}

Table 2.

Actual urbanization pattern dynamics and predicted 2020 and 2030 transitions according to the two scenarios.

Category		Scenario (Year predicted)	No. of input factors	Area of built-up land (km²)	% of built-up land (no. of pixels)
Actual transition	1990	––	––	697.3	5.87 (2,750)
	2000	––	––	1211.9	10.16 (4,760)
	2010	––	––	1561.9	13.02 (6,100)
Expected transition for scenario B	Third new-town*	––	––	34.3	0.29 (136)
Predicted transition		A (2020)	8	1965.9	16.4 (7,684)
		A (2030)	8	2037.8	17.0 (7,965)
		B (2020)	9	2733.1	22.8 (10,682)
		B (2030)	9	3176.6	26.5 (12,415)

*Third new-towns are not actual, but proposed.

Kappa: 0.998, Accuracy: 0.98, Precision: 0.98, Sensitivity: 0.99, Specificity: 0.74.

To minimize error level, we trained the RNN over 20,000 cycles for all models, and we produced two automated statistics—the kappa coefficient and the cell-matching rate—for every 1,000 cycles. As a result of neural network training, the training cycle with the highest match result from two categorial land-use data sets (actual urbanization pattern map and simulated map) was chosen for assessment and future prediction. The overall results indicate that four prediction outputs demonstrated statistical reliability high enough to provide acceptable predictions. Supplementary Figure 4 shows the training process for every 1,000 cycles.

Table 2 shows that the new development plans in the SMA will help accelerate urbanization. While the built-up area under Scenario A is expected to increase by 17.0% by 2030, a larger 26.5% increase is expected under Scenario B with the new-town developments and new transit stations. It can be inferred from the difference that the new development plan is expected to accelerate urbanization in the SMA and, thus, hamper efforts to create geographically balanced urban growth nationwide.

Figures 4 and 5 show the prediction outputs under Scenarios A and B for 2020 and 2030. As shown in Figure 4, urbanization occurs relatively evenly around Seoul by 2020 and 2030 under Scenario A. However, Scenario B suggests different prediction outputs. Since input factors for 2020 do not exist, the prediction outcomes for 2020 and 2030 were independently produced based on 1990–2010 historical input data. Thus, some 2030 development pattern prediction areas do not include the 2020 projection, and a comparison map of the prediction results between 2020 and 2030 was created as shown in Figure 6. Once the new development plans are considered, urbanization will be concentrated in the area along the eastern border of Seoul to the northeast of the Gyeonggi province by 2020 and 2030. One possible explanation for this difference is the location of Wangsuk (⑥), one of the six new-town projects planned outside of Seoul’s northeastern border. It occupies 35% of the total new-town project area and will accommodate the largest population size: 66,000. Thus, the ripple effect from the development in this area is expected to be the greatest, and the northeast area is expected to see the most substantial increase in urbanized area. This indicates that the new-town projects will intensify the population concentration in the SMA.

Figure 4.

A possible 2020 (left) and 2030 (right) urbanization pattern according to Scenario A.

Figure 5.

A possible 2020 (left) and 2030 (right) urbanization pattern according to Scenario B.

Figure 6.

Comparison of prediction results between Scenarios A and B in 2020 (left) and 2030 (right).

Urbanization determinants’ influence

To quantify each input factor’s influence, this study utilized the influence test approach developed by Pijanowski, Shellito, Bauer, and Sawaya (Pijanowski et al., 2001). By removing one variable and repeating the training process, we generated eight (Scenario A) and nine (Scenario B) different accuracy results for each removed variable model. We then compared the statistical results with the full model outputs to determine whether the variable model produced a higher accuracy value than the overall model. The dropped factor is not expected to exert a significant effect in terms of increasing urbanized areas in the SMA.

Overall, all factors within both scenarios turned out to be sound drivers in urbanization because all the dropped variables elicited lower output statistics than those included in the full model. Table 3 shows the dropped variables, the normalized influence, rank, and accuracy of each variable according to their influence (1 = weakest; 8 = strongest), and the difference in effect between the two scenarios. Since GTX variable was only used to predict the urbanization in scenario B, the variable was excluded to compare the influence of variable between scenario A and B. When the employment rate and homeownership dropped, the model’s accuracy was the lowest, that is, these factors influenced the model performance more significantly than the other factors. Surprisingly, all three transportation- and accessibility-related factors turned out to be not as influential as expected in both scenarios. Considering that most rail transit (including KTX and GTX) users and highway users travel relatively long distances, accessibility-related variables seemed to exert only marginal influence. These factors may be more influential when expanding the study area to a whole country. Furthermore, population density proved to be a primary factor in predicting urbanization in Scenario B, but this driver exerted a positive, but weak, influence in Scenario A. This partially may be because Scenario B reflected significant population growth expected from the new development projects. However, its influence seemed to be relatively low in Scenario A, as current population migration remained stable.

Table 3.

Differences in variable influence between the two scenarios.

Variable	Scenario A		Scenario B		Diff (1)–(2)
Variable	Influence (1)	Rank	Influence (2)	Rank	Diff (1)–(2)
KTX	0.972	1	0.966	1	0
GTX	––	––	0.966	––	––
Highway	0.920	2	0.908	2	0
Secondary industry	0.884	4	0.899	3	−0.143
Working population	0.880	5	0.869	5	0
Employment	0.430	8	0.850	7	−0.143
Population density	0.907	3	0.841	6	+0.429
Home ownership	0.706	7	0.793	8	+0.143
High school graduate	0.860	6	0.883	4	−0.286
Full model	0.977		0.977

Discussion

This study strived to predict future urbanization transitions in the SMA and to identify this urbanization’s influential driving factors. Future urbanization in the SMA was projected as a case study. The Korean government has worked hard for the past four decades to seek balanced nationwide urban growth and minimize growing socioeconomic inequality between the SMA and the rest of the country, but overpopulation and housing market instability in the SMA continue to grow. The first step in tackling this population concentration issue is to enhance the ability to predict future urbanization patterns. Accurate data-based spatial prediction modeling will help policy makers develop the best strategies for the future.

This research found several interesting trends. First, the accuracy outputs demonstrated that the model has high statistical reliability—enough for its predictions to be acceptable in diverse conditions and scenarios. To minimize error level and obtain the best output, we trained each model for 20,000 cycles and calculated two different statistics to compare the actual and predicted built-up area land cover pattern transitions. The kappa statistics and cell-matching rate were recorded at over 98% after 20,000 cycles. Considering that 60–80% accuracy generally is viewed as a highly acceptable level, the LSTM-based model proved to be capable of determining the areas with a high risk of rapid urbanization in the future.

Second, the government-driven, large-scale development projects are expected to accelerate urbanization, increasing the built-up areas within the SMA. The modeling result predicts that Scenario B (with the six new-town developments and three GTX projects to be considered) will lead to 55.8% more built-up areas by 2030 than Scenario A. Considering that more built-up areas often mean a higher population concentration, the government probably needs to establish planning strategies in response to the rise in additional housing and infrastructure demand, which go against its own principle of decentralizing the population and balancing nationwide urban growth. This finding may convince the government to reconsider the proposed large-scale new-town projects for long-term sustainable goals.

Third, the variable influence test results indicate that all the variables used in this research may help accelerate urbanization. Among them, homeownership and employment rates exert a stronger influence than the other factors in Scenarios A and B. Homeownership is associated highly with personal wealth and job availability. Considering that people and jobs follow each other, solutions to the urbanization issue should align closely with housing and employment conditions. An increase in high-quality housing supply in the suburbs and employment dispersal policies can promote decentralization of Seoul’s population, yet simultaneously reinforce population concentration in the SMA. Unlike previous research, we found that accessibility-related variables were less influential. Considering that highways and railways generally are used for long-distance travel, these factors may be more prominent when expanding the study area to a whole country. Population density was one of the most influential variables in Scenario B, but surprisingly, population density played a less significant role in Scenario A. Considering that the study area’s current demographic transformation and urbanization pattern is stable, population density might not help stimulate urbanization. However, once the population growth (expected from the new-town developments) is reflected in Scenario B, population density may be more influential. The LSTM-based prediction model has been demonstrated to be a potentially good alternative for predicting future land-use pattern transitions by improving both model performance and reliability.

Based on the modeling results, it is possible that the outer area of Seoul’s northeastern border would be heavily urbanized by 2030 under Scenario B (with the proposed new-town developments). The government may need to prepare in advance for the increased demand for houses and infrastructure in the expected built-up areas on the outskirts (approximately within 20–25 km of the center of Seoul). According to the influence test, increasing homeownership and employment rates may be significant and influential factors, and it can be inferred that housing- and economic growth–related policies would help disperse the city population to the outskirts, although this also could intensify the population concentration within the SMA and widen the gap between the SMA and the rest of the country.

While the LSTM-based model has been demonstrated to predict future spatial transitions and quantify variable influences effectively (particularly when a thorough and clear land-use inventory is available), this study has limitations. First, the prediction model requires excessively long training times to stabilize the error level and obtain the best output in the neural network process. For this research, the training processes took over 3 h for each of 21 different models, including four full models (for two target years under two scenarios) and 17 variable influence models (eight for Scenario A and nine for Scenario B). Second, although the influence test yields quantified results that help decide which variables are more influential than others, the chosen methods were unable to measure the influence level. Qualitative or noisy input variables could create a problem when dealing with highly correlated data. Last, although LSTM’s predictive power was proved to be highly acceptable, with 98% accuracy, this research failed to compare the model with other prediction models directly.

Conclusions

Overall, this study sought to predict future urbanization trends under two different scenarios and identify contributing factors’ influence, thereby providing guidance for sustainable urban growth policies. Land-use/land cover (LUCC) models typically have been applied to predict large-scale land cover change, but most analyses have been done with a lack of accuracy in testing. LSTM-based predictions would be a useful alternative to find areas at greater risk of urbanization if a thorough and clear input pattern and factor data are available. This study’s significant findings are summarized as follows: (1) The new development projects in the SMA are expected to accelerate urbanization about 1.5 times faster compared with the scenario without the proposed projects, with the LSTM-based model effective in predicting spatial pattern dynamics, and (2) socioeconomic variables exert substantially more influence than accessibility- and transportation-related variables.

While several studies have been conducted to predict future possible development patterns based on historical pattern dynamics, this study develops a more accurate deep-learning–based prediction model to quantify and visualize the magnitude of the future impact expected from government-driven new-town developments in two different scenarios. The research outcomes will contribute to quantifying the impacts of new developments to make spatial pattern predictions with accurately visualized land cover transformation maps. Based on appropriate data source and qualified planning theories, this research has demonstrated that how urban development patterns in Seoul are likely to change, providing the evidentiary material necessary to establish long-term urban growth policies. Further research will be necessary to develop a user-friendly modeling interface for local planners and policy makers who are not familiar with complicated deep-learning techniques.

Supplemental Material

Supplemental Material - Forecasting the urbanization dynamics in the Seoul metropolitan area using a long short-term memory–based model

Supplemental Material for Forecasting the urbanization dynamics in the Seoul metropolitan area using a long short-term memory–based model by Changyeon Lee, Jaekyung Lee, and Sungjin Park in Environment and Planning B: Urban Analytics and City Science

Footnotes

Acknowledgements

The authors would like to thank the anonymous reviewers for their helpful comments.

Author contributions

JL outlined the methodology, conducted the analysis, pre-processed data, and wrote the manuscript. CL substantially contributed to the design of the study, provided key suggestions for improving the methods, and edited the manuscript. SP rewrote abstracts and introduction and revised the whole manuscript.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2019S1A5A8032562).

Supplemental Material

Supplemental material for this article is available online.

Author Biographies

Changyeon Lee: Dr. Lee is currently an assistant professor at the Department of Urban Engineering, Jeonbuk National University in South Korea. He completed his master's degree in city and regional planning at the University of Pennsylvania and obtained a Ph.D. degree in urban and regional planning at the University of Florida. His research interests focus on land use and environmental planning using spatial data and quantitative methods.

Jakyung Lee: Dr. Lee is currently an assistant professor at Department of Urban Design and Planning, Hongik University in South Korea. He completed his master degree in urban planning, in 2011 at University of Pennsylvania and obtained his Ph.D in urban and regional science at Texas A&M University. His areas of interests include land use planning, spatial big data analysis and deep learning modeling.

Sungjin Park: Dr. Park is currently an Associate Professor in the Department of Urban Design and Planning, Hongik University in Seoul, Korea. He completed his Master degree in City and Regional Planning in 2001 at University of California, Berkeley and earned his Ph.D. in City and Regional Planning in 2008 at University of California, Berkeley. His areas of interest include walkability, sustainable transportation, and smart growth.

References

Andersson

Quigley

Wilhelmsson

(2009) Urbanization, productivity, and innovation: Evidence from investment in higher education. Journal of Urban Economics 66: 2–15.

Antrop

(2004) Landscape change and the urbanization process in Europe. Landscape and Urban Planning 67: 9–26.

Arvin

Pradhan

Norman

(2015) Transportation intensity, urbanization, economic growth, and CO2 emissions in the G-20 countries. Utilities Policy 35: 50–66.

Atack

Bateman

Haines

, et al. (2010) Did railroads induce or follow economic growth?: Urbanization and population growth in the American Midwest, 1850–1860. Social Science History 34: 171–197.

Azari

Papapetrou

Denic

, et al. (2019) Cellular traffic prediction and classification: A comparative evaluation of LSTM and ARIMA. In International Conference on Discovery Science. Springer.

Batty

Xie

Sun

(1999) Modeling urban dynamics through GIS-based cellular automata. Computers, Environment and Urban Systems 23: 205–233.

Buhaug

Urdal

(2013) An urbanization bomb? Population growth and social disorder in cities. Global Environmental Change 23: 1–10.

Hong

(2011) Testing urbanization economies in manufacturing industries: urban diversity or urban size? Journal of Regional Science 51: 585–603.

Grant

(2012) Urbanization and the Employment Opportunities of Youth in Developing Countries. Background paper prepared for EFA Global Monitoring Report.

10.

Grekousis

Manetos

Photis

(2013) Modeling urban evolution using neural networks, fuzzy logic and GIS: The case of the Athens metropolitan area. Cities 30: 193–203.

11.

S-K

(2010) Housing, social capital and community development in Seoul. Cities 27: S35–S42.

12.

Henderson

(2002) Urbanization in developing countries. The World Bank Research Observer 17: 89–112.

13.

Hochreiter

(1998) The vanishing gradient problem during learning recurrent neural nets and problem solutions. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems 6: 107–116.

14.

Jones

(1991) Urbanization issues in the Asian-Pacific Region. Asian-Pacific Economic Literature 5: 5–33.

15.

Kim

C-H

Kim

K-H

(2000) The political economy of Korean government policies on real estate. Urban Studies 37: 1157–1169.

16.

Kim

C-H

Park

Y-H

Park

I-K

(2003) The Characteristics and Policy Implications of Territorial Structure Changes in Korea. Anyang: Korea Research Institute for Human Settlements.

17.

Kim

Han

(2012) Seoul. Cities 29: 142–154.

18.

Kotavaara

Antikainen

Rusanen

(2011) Population change and accessibility by road and rail networks: GIS and statistical approach to Finland 1970–2007. Journal of Transport Geography 19: 926–935.

19.

Kumar

Kober

(2012) Urbanization, human capital, and cross-country productivity differences. Economics Letters 117: 14–17.

20.

Liang

Liu

, et al. (2018) Delineating multi-scenario urban growth boundaries with a CA-based FLUS model and morphological method. Landscape and Urban Planning 177: 47–63.

21.

Liddle

Messinis

(2015) Which comes first–urbanization or economic growth? Evidence from heterogeneous panel causality tests. Applied Economics Letters 22: 349–355.

22.

Liu

Xiao

, et al. (2021) Simulation of Dynamic Urban Expansion under Ecological Constraints Using a Long Short Term Memory Network Model and Cellular Automata. Remote Sensing 13: 1499.

23.

Newman

Lee

Berke

(2016) Using the land transformation model to forecast vacant land. Journal of Land Use Science 11: 450–475.

24.

Nghiep

(2001) Predicting housing value: A comparison of multiple regression analysis and artificial neural networks. Journal of Real Estate Research 22: 313–336.

25.

Pijanowski

Shellito

Bauer

, et al. (2001) Using GIS, artificial neural networks and remote sensing to model urban change in the Minneapolis–St. Paul and Detroit Metropolitan areas. In Proceedings, American Society of Photogrammetry and Remote Sensing Annual Conference, April 23–27, 2001, St. Louis, MI.

26.

Pijanowski

Brown

Shellito

, et al. (2002) Using neural networks and GIS to forecast land use changes: a land transformation model. Computers, Environment and Urban Systems 26: 553–575.

27.

Pontius

Boersma

Castella

J-C

, et al. (2008) Comparing the input, output, and validation maps for several models of land change. The Annals of Regional Science 42: 11–37.

28.

Sato

Zenou

(2015) How urbanization affect employment and social interactions. European Economic Review 75: 131–155.

29.

Shahbaz

Lean

(2012) Does financial development increase energy consumption? The role of industrialization and urbanization in Tunisia. Energy Policy 40: 473–479.

30.

Squartini

Hussain

Piazza

(2003) Preprocessing based solution for the vanishing gradient problem in recurrent neural networks. In: Proceedings of the 2003 International Symposium on Circuits and Systems. IEEE.

31.

Tayyebi

Pekin

Pijanowski

, et al. (2013) Hierarchical modeling of urban growth across the conterminous USA: developing meso-scale quantity drivers for the Land Transformation Model. Journal of Land Use Science 8: 422–442.

32.

Turner

Meyer

(1991) Land use and land cover in global environmental change: considerations for study. International Social Science Journal 43: 669–679.

33.

Vatuk

Coleman

(1972) Kinship and Urbanization: White Collar Migrants in North India. Berkeley, CA: University of California Press.

34.

Verburg

Kok

Pontius

et al. (2006) Modeling Land-Use and Land-Cover Change. Land-Use and Land-Cover Change. New York: Springer, 117–135.

35.

Waddell

(2002) UrbanSim: modeling urban development for land use, transportation, and environmental planning. Journal of the American Planning Association 68: 297–314.

36.

Wang

Liu

Sun

, et al. (2016) Accessibility impact of the present and future high-speed rail network: a case study of Jiangsu Province, China. Journal of Transport Geography 54: 161–172.

37.

Yoon

(2019) A study on the direction of the third phase new town development in Seoul metropolitan area through expert survey method. LHI Journal of Land, Housing, and Urban Affairs 10: 43–55.

38.

Zhang

Y-J

Liu

Zhang

, et al. (2014) The impact of economic growth, industrial structure and urbanization on carbon emission intensity in China. Natural Hazards 73: 579–595.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.81 MB