The Spatial–Temporal Characteristics of Land De-Urbanization in Metropolises: A Case Study of the Guangdong–Hong Kong–Macao Greater Bay Area

Abstract

With a series of redevelopment activities, such as land consolidation and urban renewal, many cities in China have experienced land de-urbanization phenomena. These include the conversion of construction land into green spaces (such as parks, forests, and lawns), blue spaces (such as rivers, lakes, and wetlands), and farmland. However, there is currently limited research on diverse land de-urbanization types and pathways. This study focuses on investigating the land de-urbanization in the Guangdong–Hong Kong–Macao Greater Bay Area (GBA) from 2014 to September 2023 using the Continuous Change Detection and Classification (CCDC) method. The results demonstrate that the GBA experienced 72.74 square kilometers of de-urbanization during the study period, primarily through the conversion of construction land to land with low plant coverage, including grassland and farmland. There were significant differences in the quantity and spatial agglomeration of de-urbanization between cities and within individual cities. Temporally, de-urbanization predominantly occurred in the period of 2016 to 2021, with a sharp decline in 2022. The temporal changes were significantly influenced by urban renewal policies and the impact of the COVID-19 pandemic. In terms of spatial clustering characteristics, the de-urbanization process in the GBA exhibited spatial agglomeration but was primarily characterized by low-level clustering. This study also examines the correlations between de-urbanization and factors including location and the stage of urbanization. The analysis showed that de-urbanization within cities tended to concentrate near the main urban roads within a range of 10–30 km from city centers. The trend of de-urbanization followed a pattern that is consistent with the Northam curve, where de-urbanization tends to increase during the rapid urbanization phase and decline as urbanization reaches a mature stage. Overall, this study provides valuable insights for the redevelopment of construction land within the context of ecological civilization construction. It also offers suggestions for urban land development and redevelopment in metropolitan areas.

1. Introduction

Since the late 20th century, urbanization has undergone rapid development globally. According to a survey by the United Nations Department of Economic and Social Affairs [1], the global urban population reached 4.2 billion in 2018, accounting for 55% of the total population. Rapid urbanization has brought about drastic changes in land use in metropolitan areas and has led to a multitude of ecological and social challenges, such as environmental pollution, farmland loss, urban heat island effect, urban flooding, inadequate public facilities, and traffic congestion [2,3,4,5]. To address these challenges, various regions have implemented a series of land management measures, such as urban renewal, land consolidation, or ecological restoration, involving the restructuring of land space and regional functionality. This process entails the adjustment, renovation, and reconstruction of improperly utilized structures and land [6]. Some construction land is transformed into green spaces (such as parks, forests, and lawns), blue spaces (such as rivers, lakes, and wetlands), and farmland. This phenomenon refers to the reverse transformation of construction land into non-construction land, known as land de-urbanization [7].

In the past, land de-urbanization was not widely discussed due to the perception that construction land was irreversible. Many studies on land use change, particularly those focusing on urban expansion, often employ unidirectional hypothesis testing to exclude the possibility of reverse changes in construction land use [8,9]. However, based on the current domestic and international urban development trends, phenomena such as the regreening of brownfields, the restoration of water bodies, the ecological restoration of industrial and mining sites, and the redevelopment of demolished land have become increasingly common and significant. According to the “Special Plan for Urban Renewal in Beijing (Urban Renewal Plan for the 14th Five-Year Period in Beijing)”, the ecological conservation area of Beijing has the potential for approximately 12 million square meters of building renewal. In this area, existing construction land, such as old factories and industrial parks, will be transformed into ecological spaces [10].

De-urbanization has a positive impact on environmental improvement, as it can regulate microclimate and hydrology, mitigate the urban heat island effect, and enhance the value of ecosystem services [11,12,13,14]. Simultaneously, it is significant for the improvement of residents’ quality of life [15,16]. Consequently, de-urbanization, as an ecologically oriented redevelopment approach, has gained increasing attention in recent years.

The existing research related to land de-urbanization mainly encompasses the following three aspects: (1) the monitoring of de-urbanization; (2) an analysis of the causes of de-urbanization; and (3) the impacts of de-urbanization.

The research on de-urbanization monitoring can be divided into two categories: studies on the reduction in impervious surfaces in cities and urban re-greening studies. Impervious surfaces are important indicators for measuring urbanization [17], and they have been extensively and deeply studied for a long time. They have also become important visual indicators in de-urbanization research. For instance, Deng and Zhu developed the Continuous Subpixel Monitoring Method, which can detect urban demolition and redevelopment according to the loss of impervious surface percentage (ISP) [18]. Fu also used the ISP dynamic to examine both urbanization and de-urbanization, and the results showed that the annual de-urbanization rate in Guangzhou from 2000 to 2010 ranged from 1% to 5% [7]. In urban regreening research, new indices like the normalized urban areas composite index [19] and the Regional Greenness Dynamic Index [20] have been proposed to detect green regrowth within this urbanized extent. Xia and Zhang analyzed redevelopment and the improvement of urban green coverage using the shape-weighted landscape evolution index [21].

However, it should be noted that reverse changes in construction land use can exhibit diverse pathways, and solely monitoring the reduction in impervious surfaces and regreening may not provide a comprehensive description of these changes.

The current research on the causes of urbanization is mainly focused on economic and policy aspects. For instance, Wu (2021) examined the variations in green spaces across 107 cities in China and found that economic growth has a negative impact on urban green spaces on an overall city scale [20]. But their study reveals that in older and more developed areas of cities, such economic growth actually stimulates regreening. In other words, as urban economies become more prosperous, the level of green space optimization in long-established areas tends to increase. Another study conducted by Zhang investigated the trend of regreening in the vegetation cover changes in the Yangtze River Delta region and revealed that the older urban areas in cities with higher levels of urbanization exhibit the highest inclination towards greening [22]. This may be attributed to the implementation of greening policies and urban renewal strategies by local governments in regions characterized by higher economic levels and urbanization rates, thus fostering ecological restoration. Many studies suggest that government policies play a significant role in reversing the transformation of construction land. For example, the “Balance of Arable Land” policy has promoted the transformation of construction land to arable land [23], while ecological restoration and greening policies have led to an increase in ecological land [24]. Government officials and urban planners are increasingly realizing the necessity to enhance landscape diversity by incorporating parks, playgrounds, greenways, and other open spaces through redevelopment [25].

There are extensive studies in the literature that explore the potential impacts of reverse changes in construction land use on various aspects and domains, such as environmental quality, land surface temperature, surface runoff, and human habitat quality. Researchers have primarily focused on the environmental benefits that green-oriented redevelopment can bring to cities. For instance, Zheng et al. [14] and Wu, P. et al. [26] found that reducing building density and increasing vegetation coverage during urban renewal can lower surface temperatures and alleviate the urban heat island effect. Zhou [27] pointed out the positive role of large-scale afforestation in Guangdong Province in regulating seasonal river flow. Scholars and urban planners have long advocated for integrating the concepts of renaturing and rewilding into urban planning to restore natural habitats and promote sustainable development [28,29]. Furthermore, multiple studies have shown that urban greening can enhance residents’ quality of life and well-being [30,31,32]. A study by De Soda demonstrated that brownfield redevelopment provides valuable opportunities for increasing green spaces in urban areas, resulting in benefits such as improved soil quality, habitat creation, and community economic revitalization [33].

Although there have been many studies related to de-urbanization, there are still some limitations. Firstly, the existing studies focus on regreening, and there is a lack of research on diversified de-urbanization. Different paths of de-urbanization reflect different patterns of land use structure optimization, which have varying impacts on the ecological environment and socio-economic development. Therefore, it is necessary to identify and explore diversified de-urbanization in detail. Secondly, the current research lacks a macroscopic understanding of de-urbanization. The spatio-temporal characteristics of de-urbanization and their correlations need further investigation. The results of such investigations can provide more effective guidance and decision support for urban redevelopment, which is crucial for achieving sustainable and livable urban environments.

By using the Guangdong–Hong Kong–Macao Greater Bay Area (referred to as the GBA or Greater Bay Area) as the research area, this study identifies de-urbanization from 2014 to September 2023, including the conversion of construction land into forest land, land with low plant coverage (including grassland and farmland), bare land, and water bodies. The spatial distribution and temporal information of these changes are obtained through a time series analysis of remote sensing imagery. This study analyzes the spatio-temporal differences and spatial clustering characteristics of de-urbanization on both the regional and urban scales within the Greater Bay Area. Additionally, it explores the correlation between de-urbanization and location, urban development stages, and policies.

2. Methodology

The framework of this study mainly consists of the following steps: (1) identification of de-urbanization; (2) spatial–temporal analysis of de-urbanization; and (3) correlation analysis (e.g., urban development stage, location, and policy implementation) (Figure 1). The workflow is illustrated in Figure 1.

2.1. Data Source

The data used in this study consist of remote sensing images, land use/land cover reference data, and basic geo-data.

Table 1. Data sources.

Data Type	Datasets	Resolution	Time	Source
Remote Sensing Images	USGS Landsat 8 Level 2, Collection 2, Tier 1	30 m	2014–2023.09	https://developers.google.com/earth-engine/datasets/catalog/landsat-8, (accessed on 25 September 2023)
Land Use/Land Cover Reference Data	ESRI—Sentinel-2 Land Use/Land Cover Timeseries dataset	10 m	2017–2021	https://esri.maps.arcgis.com (accessed on 20 September 2023)
	GAIA (Global Artificial Impervious Areas data)	30 m	2014	http://data.starcloud.pcl.ac.cn/zh/resource/13 [34] (accessed on 20 September 2023)
	Land use survey data	10 m	2014	Bureau of Land and Resources
Basic Geo-Data	Road	/	2023	Open Street Map
	City center	/	2023	Baidu Map

presents the data sources utilized in this study. Remote sensing images and land use/land cover reference data are used for the identification of de-urbanization areas. Time series of the bands in Landsat 8 imagery and computed indices such as the NDVI are important inputs for the CCDC model. Samples of land use classifications were taken from the ESRI land cover dataset. The time series of impervious surface data from GAIA can be used to calculate urbanization rates. The impervious surface percentages in the year 2014, extracted from GAIA, combined with land use survey data, were employed to delineate the initial extent of stock of construction land in this study. In addition, roads and urban center points were utilized for a descriptive analysis of the spatial characteristics of de-urbanization and a quantitative assessment of their relevance through a Euclidean distance analysis.

2.2. Identification and Analysis of Spatio-Temporal Distribution of De-Urbanization

This section focuses on methods for identifying de-urbanization areas and the results of the spatial autocorrelation analysis.

2.2.1. Identification of Land De-Urbanization

This study utilizes remote sensing time series data and the CCDC (Continuous Change Detection and Classification) model [35] to identify de-urbanization areas. Before running the CCDC, it is important to determine the land use classification system and develop identification rules for de-urbanization.

Land Use Classification Systems

In this section, hypotheses on the potential pathways of de-urbanization are firstly presented. Drawing upon measures of land consolidation and urban renewal, we consider several possible land de-urbanization pathways, including (1) the conversion of construction land into forest land, (2) the conversion of construction land into grassland, (3) the conversion of construction land into agricultural land, (4) the conversion of construction land into water bodies, and (5) the conversion of construction land into bare land. This study extensively considers multiple existing land use classification systems. In accordance with the five potential pathways and by drawing upon the Local Climate Zone (LCZ) classifications adopted by Zhu [36], grassland and agricultural land are merged into a single category of “Low plants”. The final determined land use classification system and corresponding land descriptions are presented in

Table 2. Land use classification system used in this study and their descriptions.

	Category	Description
1	Construction land	Artificial cover and structures such as pavement, concrete, brick, stone, and other humanmade impenetrable cover types.
2	Forest	Dense or scarce forest.
3	Low plants	Grassland, farmland, etc.
4	Bare land	Land including bare rock and bare soil.
5	Water	Lake, river, pool, reservoir, etc.

.

The CCDC Algorithm for De-Urbanization Identification

The CCDC algorithm is capable of detecting many kinds of land use/land cover changes continuously by utilizing all available remote sensing images. It generates land cover maps for any given time [35]. The algorithm primarily consists of two major steps: (1) continuous change detection and (2) continuous land cover classification.

This study applies the CCDC temporal breakpoints algorithm using the Google Earth Engine (GEE) platform. The selected breakpointBands encompass various spectral bands, including Blue, Green, Red, Near-Infrared (NIR), Shortwave Infrared 1 (SWIR1), and Shortwave Infrared 2 (SWIR2), in addition to several indices such as the Normalized Difference Vegetation Index (NDVI), the Normalized Difference Built-up Index (NDBI), and the Modified Normalized Difference Water Index (MNDWI). The calculation formulas for these bands and indices are provided in

Table 3. Calculation formulas for selected indices.

Band/Index	Calculation Formula
Normalized Difference Vegetation Index (NDVI)	NDVI = $\frac{NIR-RED}{NIR+RED}$	(1)
Normalized Difference Built-Up Index (NDBI)	NDBI = $\frac{SWIR1-NIR}{SWIR1+NIR}$	(2)
Modified Normalized Difference Water Index (MNDWI)	MNDWI = $\frac{GREEN-SWIR1}{GREEN+SWIR2}$	(3)

.

(1) Continuous Change Detection: This step aims to detect change points in land surface cover, referred to as “breakpoints.” It is based on the assumption that land cover remains constant over time. A model based on sine and cosine estimation is used to predict the remote sensing observations at any given date. Discrepancies between the predicted and observed satellite data are utilized to identify changes in land cover. In this study, the minimum number of consecutive observations (minObservations) required for a pixel to be considered as exhibiting “change” is set to 6. If a pixel consistently demonstrates “change” over six consecutive instances, it can be inferred that a change in land cover has occurred. The CCDC algorithm designates this change point as a “breakpoint,” subsequently dividing the time series data after the breakpoint into a new “segment.” The outcomes of this initial step encompass information regarding breakpoints, fitting coefficients for each band within each time segment, the Root Mean Square Error (RMSE), and the band with the maximum normalized residual at the breakpoints.

(2) Continuous Land Cover Classification: In this step, the CCDC algorithm utilizes a non-traditional classification approach by employing a Random Forest classifier. It utilizes the coefficients obtained from the time series models generated in the initial step as inputs for the land cover classification process. By classifying these coefficients, the algorithm can determine the corresponding land cover types for each time series model throughout the entire observation period. The samples for land cover classification are obtained from the ESRI Land Use Reference dataset. A random sampling approach was employed to select 100 pixels for each land use/cover category. These pixels were then individually examined by comparing them with historical imagery through Google Earth Pro to ensure the reliability of the training data. Any samples whose assigned labels did not match the actual scene were manually removed. Ultimately, a total of 1382 sample points were utilized in the Random Forest classifier.

The identification method for de-urbanization in this study involves using the intersection between the pixels classified as construction land in the first segment of the CCDC and the 2014 land use classification data to determine the existing stock of construction land. Then, pixels are selected that are classified as forest or shrubland, bare land, or water in the second segment after the change.

This method provides the location information for de-urbanization, enabling the generation of maps depicting different types of change pathways. And the band “tBreak” provided by the model output offers temporal information regarding the occurrence of changes.

2.2.2. Spatial Autocorrelation Analysis of De-Urbanization

Global spatial autocorrelation describes the overall distribution of land de-urbanization, and it determines whether the de-urbanization area is spatially agglomerated or discrete. The Moran index is used in this study. Moran’s I is calculated as follows:

$$\mathrm{I}=\frac{\mathrm{n}\sum _{\mathrm{i}=1}^{\mathrm{n}}\sum _{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{W}}_{\mathrm{i}\mathrm{j}}({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})({\mathrm{x}}_{\mathrm{j}}-\overline{\mathrm{x}})}{\sum _{\mathrm{i}=1}^{\mathrm{n}}\sum _{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{W}}_{\mathrm{i}\mathrm{j}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})}^2}$$

(4)

Spatial processes and patterns can vary across different regions within a study area. A global spatial autocorrelation analysis may mask or overlook these local variations. Therefore, when there is spatial heterogeneity in the data, and different regions have distinct spatial patterns or relationships, a local spatial autocorrelation analysis would be useful. In this study, the local clustering degree of the de-urbanization area can be examined by employing the local spatial autocorrelation approach. This study utilizes the Local Indicators of Spatial Autocorrelation (LISA) method, assessing the spatial clustering of de-urbanization areas within local regions by employing the local Moran’s I and Getis-Ord ${\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}$ statistics. The local Moran’s I examines the similarity between each location and its neighboring locations, revealing local spatial patterns. A positive local Moran’s I indicates that high values are clustered around other high values or low values are clustered around other low values. Conversely, a negative local Moran’s I indicates an opposite spatial pattern, where high values are surrounded by low values or vice versa. The local Moran’s I index is calculated as follows:

$${\mathrm{I}}_{\mathrm{i}}=\frac{\mathrm{n}({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}}){\sum }_{\mathrm{j}}{\mathrm{w}}_{\mathrm{i}\mathrm{j}}({\mathrm{x}}_{\mathrm{j}}-\overline{\mathrm{x}})}{{\sum }_{\mathrm{i}}{({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})}^2}$$

(5)

where ${\mathrm{I}}_{\mathrm{i}}$ is the local Moran’s I index; $\mathrm{m}$ is the number of study objects; ${\mathrm{x}}_{\mathrm{i}}$ is the observed value; $\overline{\mathrm{x}}$ is the mean value of ${\mathrm{x}}_{\mathrm{i}}$; and ${\mathrm{w}}_{\mathrm{i}\mathrm{j}}$ is the spatial weight.

The LISA clustering map generated from the results of the local Moran’s I index can identify the locations of different patterns of spatial agglomeration or discrete phenomena in the study area, including High–High Clustering areas, Low–Low Clustering areas, High–Low Clustering areas, and Low–High Clustering areas.

The Getis-Ord ${\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}$ statistic is used to identify the distribution of local hotspots or coldspots, representing areas with significantly high or low values within a local region. The application of the Getis-Ord ${\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}$ statistic allows us to identify significant local spatial clustering areas, where high-value areas are surrounded by high-value areas or low-value areas are surrounded by low-value areas.

The Getis-Ord ${\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}$ index is calculated as follows:

$$G_i^{\mathrm{*}}=\frac{\sum _{j=1}^n{w}_{i,j}{x}_j-\overline{X}\sum _{j=1}^n{w}_{i,j}}{S\sqrt[]{\frac{\left[n\sum _{j=1}^n{w}_{i,j}^2-{(\sum _{j=1}^n{w}_{i,j})}^2\right]}{n-1}}}$$

(6)

In the analysis of the Getis-Ord ${\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}$ statistic, if the ${\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}$ score is greater than 0 and higher in value, it indicates a tighter clustering of high-attribute values (forming hotspots). Conversely, if the ${\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}$ score is less than 0 and lower in value, it indicates a tighter clustering of low-attribute values (forming coldspots).

The above spatial econometric methods were implemented using GeoDa 1.20.

2.3. Accuracy Assessment for De-Urbanization Identification

After identifying the de-urbanization pixels in the study area, accuracy validation is an essential step. In the validation, a total of 200 sample points were randomly selected from the de-urbanization extraction results, with 50 points selected from each type of de-urbanization pathway. The validation was conducted using visual interpretation in Google Earth, comparing the historical imagery from the initial year (2014) with the change year.

3. Study Area

The Greater Bay Area in China, comprising the Hong Kong Special Administrative Region, the Macau Special Administrative Region, and the cities of Guangzhou, Shenzhen, Zhuhai, Foshan, Huizhou, Dongguan, Zhongshan, Jiangmen, and Zhaoqing in Guangdong Province, is chosen as the research area (as shown in Figure 2). With a total area of 56,000 square kilometers, the Greater Bay Area is a region of significant strategic importance within the broader context of China’s national development. By 2022, its population had surpassed 86 million, positioning it as one of China’s most prosperous and economically advanced regions.

Diverging from other international bay areas and domestic urban agglomerations, the GBA exhibits a polycentric spatial structure, featuring diverse institutional mechanisms and notable differentiation across various gradients. In terms of jurisdictional area, Macao spans a mere 33 km², which is much smaller than other cities in the GBA. In contrast, the largest city, Zhaoqing, is approximately 450 times larger than Macao. Economic development unveils a distinct polarization phenomenon, which is particularly evident in the development gap between cities like Guangzhou, Shenzhen, and Hong Kong and in less-developed cities such as Zhaoqing and Jiangmen. For instance, there exists a tenfold disparity between Shenzhen, Guangzhou, and Zhaoqing. Furthermore, significant disparities exist in the industrial structure and infrastructure levels among each city.

From an overall perspective, the GBA is one of the regions in China with the highest level of urbanization development. It has a significant area of existing construction land (as shown in Figure 2, depicting the distribution of existing construction land in 2014). Moreover, the GBA has been at the forefront of urban renewal efforts in China, which were initiated at an early stage [37]. In recent years, the GBA has made significant contributions to urban ecological construction. These conditions provide ample samples and a rich context for studying de-urbanization in the region. Against this backdrop, what are the characteristics of de-urbanization among different cities? Considering the considerable estimation of the reverse changes in construction land in the GBA and the existing differences among cities in the Bay Area, it is chosen as the research region for this study.

4. Results

4.1. The Identification Results and Accuracy of De-Urbanization

4.1.1. Location and Distribution of De-Urbanization in the GBA

By applying the CCDC model to the GBA using a time series of Landsat 8 images from 2014 to September 2023, we identified the occurrence of de-urbanization for each individual stock of construction land and determined the time points at which it occurred. The results revealed approximately 72.74 km² (727,375 pixels) of de-urbanization within the GBA. When analyzing the different land cover transitions (as shown in Figure 3), it can be observed that the reverse changes in construction land primarily involve a transformation into low plants and bare land, while the presence of water and forest is relatively limited.

Table 4. Zonal statistical results of de-urbanization and the proportion of de-urbanization to stock of construction land.

Cities	De-Urbanization (km2)	Stock of Construction Land (km2)	Proportion of De-Urbanization to Stock of Construction Land (%)
Guangzhou	15.96	1609.03	0.99
Shenzhen	9.4	872.95	1.08
Foshan	11.05	1368.48	0.81
Huizhou	14.65	882.13	1.66
Jiangmen	7.41	909.27	0.81
Hongkong	1.12	184.11	0.61
Zhuhai	4.6	403.03	1.14
Zhonshan	3.36	636.50	0.53
Macao	0.0042	16.68	0.03
Dongguan	5.18	1105.70	0.47
Zhaoqing	0	282.70	0

presents the zonal statistical results of de-urbanization and the proportion of de-urbanization to existing construction land for each city in the GBA. It can be observed that there are significant variations in reverse changes in construction land among cities in the GBA. The city with the highest amount of de-urbanization is Guangzhou, where approximately 15.96 km² of existing construction land have undergone reverse changes. Significant changes are also observed in Huizhou and Foshan. When considering existing construction land, the cities with the highest proportions of reverse changes are Huizhou, Zhuhai, and Shenzhen. The relatively high proportion of reverse changes in relation to the existing construction land reflects the emphasis these cities place on the socio-ecological benefits of redevelopment activities. Macau has a minimal reverse change area of only 0.0042 km², accounting for 0.03% of the total. The detection result for Zhaoqing is 0 km², indicating that no de-urbanization was identified. The absence of de-urbanization in Zhaoqing can be attributed to its position in the economic development hierarchy within the GBA. Zhaoqing belongs to the last tier of economic development, with a relatively slow urbanization process and ample room for urban and ecological development. There is a low demand for exploiting existing construction land. Although Hong Kong and Macau have achieved high levels of urbanization and have been early adopters of urban renewal and other inventory exploration measures, the limited land area makes large-scale reverse changes in construction land unfeasible.

A fishnet with 500 m × 500 m grid units was created to calculate the density of de-urbanization pixels. The results of the regional analysis reveal that the count of pixels exhibiting de-urbanization within each grid cell ranges from 1 to 855. To effectively classify this range, we employed the natural breaks method and categorized it into five levels. A higher count of de-urbanization pixels within a grid cell signifies a greater density of reverse changes in construction land. A visual representation of the grid analysis can be observed in Figure 4. Overall, the density of de-urbanization is relatively low, with the “Very Low” level prevailing. This level is characterized by grid cells containing between 1 and 22 pixels. There are a few instances of high de-urbanization density near the “Low” and “Moderate” density grids in the western area of Shenzhen, the central–western part of Guangzhou, and southwestern parts of Huizhou. These specific areas have experienced a significant phenomenon of “de-urbanization”, characterized by large-scale demolitions. It is particularly noteworthy that in Huizhou, although the extent of de-urbanization may not be extensive on the map, the concentration of de-urbanization in the central and southwestern regions and its remarkably high density result in a substantial quantity of de-urbanization.

4.1.2. Accuracy Assessment for De-Urbanization Identification

The accuracy of the manual random verification of de-urbanization was validated by comparing it with samples extracted from historical high-resolution imagery, resulting in an overall accuracy rate of 79%. The accuracy rates for the specific pathways are as follows: the accuracy for the pathway transitioning to forest land is 82%, the accuracy for the pathway transitioning to low plants is 88%, the accuracy for the pathway transitioning to bare land is 78%, and the accuracy for the pathway transitioning to water is 66%. Figure 5 illustrates three typical examples of the detection results of de-urbanization. From left to right, the images depict the initial existing construction land landscape in 2014 obtained from Google Earth, the landscape of the changing year, and the overlay of de-urbanization pixels detected by the CCDC overlay with the NDVI image, the NDVI Time Series. The pixels within the red boundary represent the detection results for de-urbanization.

4.2. Temporal Analysis

The “tBreak” in the CCDC detection results reflects the dates of detected land cover changes. The statistical analysis of the change dates for different land cover types is illustrated in Figure 6. It is evident that there is a significant fluctuation in the number of pixels associated with different de-urbanization pathways. The de-urbanization path of low plants reached their peak in 2016, followed by a decline, until reaching relatively higher levels again in 2019 and 2021. The number of de-urbanization paths for bare land exhibited an overall increasing trend until starting to decrease in 2022. There were relatively fewer de-urbanization path towards forests and water bodies, but it can be observed that these two categories achieved higher values in 2016 and 2019. Overall, 2016, 2019, and 2021 were peak periods for reverse changes in construction land. This is closely related to urban renewal and urban greening policy measures. In 2016, cities such as Guangzhou, Shenzhen, and Foshan launched urban renewal plans. In 2019, China entered a rapid development phase of urban renewal, with central and local governments issuing consecutive policies to accelerate the transformation of old urban areas. The year of 2019 was therefore referred to as the “Year of Urban Renewal.” Additionally, at the end of 2018, Guangdong Province initiated a large-scale three-year greening action plan. In 2021, the Guangdong Provincial Government issued opinions on the implementation of scientific greening, which emphasized combining urban renewal with measures such as demolishing illegal structures and planting greenery, increasing green spaces through vacant land and vertical greening, etc., encouraging efforts to increase green spaces through land use optimization and intensifying greening efforts. A series of policies and measures have promoted the development of reverse changes in construction land.

However, in 2022, there was a sharp decline in the number of reverse changes in construction land. We believe that this was influenced by multiple factors. In August 2021, the Ministry of Housing and Urban-Rural Development issued a notice regarding the prevention of excessive demolition and reconstruction in the implementation of urban renewal actions, emphasizing strict control over large-scale demolitions of existing buildings. In March 2022, the Guangdong Provincial People’s Government issued a notice on the implementation of the key tasks and division of work for the provincial government’s work report for 2022, which mentioned promoting urban renewal through meticulous efforts and adopting more micro-reconstruction methods. These policies led to a significant decrease in the number of reverse changes in construction land.

Furthermore, the sharp decline in data in 2022 was also influenced, to some extent, by the pandemic. In 2022, many areas in Guangdong faced the most complex and challenging situations in the three-year battle against the pandemic. The declaration, approval, and implementation of urban renewal projects were hindered, which consequently affected the progress of de-urbanization.

It should be noted that the “low point” in the data for 2023 is mainly due to the monitoring data only covering the period until September 2023, resulting in a relatively short monitoring period. Based on the observations from the first eight months of 2023, the reverse changes have not shown significant recovery. However, in the long run, it is expected that reverse changes in construction land will gradually rebound.

We further associate the above results with different cities. Figure 7 presents the annual flow of existing construction land in each city towards different land categories. As shown in the figure, the main reverse change paths for construction land in Guangzhou, Huizhou, and Hong Kong are towards bare land and low plants. The de-urbanization paths in Dongguan, Foshan, Zhongshan, and Zhuhai are mainly progressing towards low plants. The reverse change paths in Shenzhen appear to be more diverse. The largest influx of bare land comes from Guangzhou due to significant conversions in the 2019–2021 period. Forest received the highest influx in 2016, primarily from existing construction land in Guangzhou and Shenzhen. The significant contributions to the low plants path can be attributed to the influx in Foshan in 2015 and 2016. Water mainly originates from the influx in Shenzhen in 2016 and 2017.

4.3. Spatial Autocorrelation Analysis

To investigate the clustering characteristics of de-urbanization, a spatial autocorrelation analysis was conducted. It is based on the fishnet-grid-based statistical result of de-urbanization in this study. Figure 8 displays the significance of spatial clustering in the grid-based density of de-urbanization using the Getis-Ord ${\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}$ statistic. Figure 9 presents the LISA cluster map for the grid-based density of de-urbanization.

First, the global Moran’s I index was used to obtain the Moran’s I value of 0.307, with a p-value of 0.001 and a z-value of 78.6236. These results indicate a statistically significant positive spatial correlation in the intensity of de-urbanization within the GBA at a confidence level of at least 99%.

According to the figure, highly clustered hotspots of de-urbanization are primarily located in the western part of Shenzhen, the central and southern regions of Huizhou, the northwestern part of Guangzhou, the northern part of Jiangmen, and Zhuhai City. The distribution in Foshan is more scattered. Some level of clustering can also be observed in Dongguan, Zhongshan, and Hong Kong. Based on the LISA cluster map for local spatial autocorrelation, there are more instances of high–high and low–low clusters compared to high–low and low–high outliers. This suggests that areas with high (low) levels of reverse changes in construction land tend to cluster spatially. The most significant pattern observed is high–high clustering, while low–low clustering exhibits a more dispersed distribution.

The above results reflect the significant influence of regional development levels and government strategies on de-urbanization. Taking Guangzhou as an example, the existing construction land in the city has undergone substantial reverse changes. However, highly concentrated reverse changes, particularly those involving larger-scale redevelopment activities, are primarily concentrated in the Huadu, Baiyun, Huangpu, and Conghua districts. The central urban area exhibits s lower spatial clustering of de-urbanization, possibly due to limited land resources, high economic development levels, and high land rent, making large-scale de-urbanization challenging. Furthermore, the introduction of the “Guangzhou Urban Renewal Measures” in 2016 emphasizes micro-redevelopment, gradually replacing the old model of large-scale demolition and construction. As a result, the number of large-scale urban renewal projects in areas with rapid urbanization has gradually decreased. In contrast, less-developed districts have continued to undertake numerous tasks related to large-scale urban renewal in recent years. On the other hand, in Huizhou City, a significant portion of the reverse changes is driven by large-scale demolition and reconstruction projects. The areas with highly concentrated reverse changes are mainly located in the Huicheng and Huiyang districts, which have higher levels of urbanization.

4.4. Correlation Analysis of De-Urbanization

Based on the previous analysis, de-urbanization is influenced by certain factors, including urban land area, levels of economic development, urban renewal policies, and the impacts of pandemics. In this section, we focus on analyzing the correlation between de-urbanization and location, as well as the development stage of urbanization.

4.4.1. Location Correlation

The distance from the city center and the proximity to major urban roads are vital indicators for measuring location. By examining the characteristics of de-urbanization based on these factors, we present the findings in Figure 10 and Figure 11.

Figure 10 illustrates that de-urbanization primarily occurs within a distance of 10–30 km from the city center. And grid cells with high densities of de-urbanization are mainly concentrated within a broader range of 10–50 km. This reflects that de-urbanization is less prevalent in the core area. Figure 11 demonstrates a concentration of de-urbanization in close proximity to primary urban roads. As the distance from these main roads increases, both the quantity and density of de-urbanization within the grid cells diminish.

4.4.2. Development Stage of Urbanization

This study has plotted the urbanization rate curve, which illustrates the changes in the level of urbanization development over time. The urbanization rate is determined based on the annual variations in the proportion of construction land to the total urban land area. Figure 12 displays the overlay of the fitted curve of urbanization rate and the stacked bar chart representing the de-urbanization trend from 2014 to September 2023. The subsequent analysis of urbanization development stages in this study excludes Hong Kong and Macau due to their unique development patterns. Based on the Northam curve [38] and the actual development of each city, this study categorizes the development stages of the cities as follows: Zhaoqing is in the initial stage, while Jiangmen, Zhongshan, Huizhou, Zhuhai, and Guangzhou are in the acceleration stage. Dongguan, Shenzhen, and Foshan are in the terminal stage. The rationale behind classifying Guangzhou in the acceleration stage, and its low urbanization rate, is attributed to the ongoing expansion of its administrative boundaries. The jurisdictional area of Guangzhou has undergone multiple adjustments, resulting in continuous expansion. The inclusion of Conghua, a newly added district, has introduced a substantial amount of undeveloped land into the city. Consequently, the overall urbanization rate in Guangzhou is not particularly high. By examining the land urbanization development curve, it becomes evident that Guangzhou is currently in the acceleration stage of urbanization.

The stacked bar chart represents the annual variation in de-urbanization quantities for each city. In the initial stage of urbanization, there is no de-urbanization observed. Excluding the special years of 2022 and 2023, cities experiencing rapid urbanization development, such as Guangzhou and Zhuhai, demonstrate an increasing trend in deurbanization quantities. Conversely, cities that have reached a stable and mature stage of urbanization development, such as Foshan, Shenzhen, and Dongguan, exhibit a decreasing trend in deurbanization quantities. This observation highlights the significant impact of the urbanization development stage on deurbanization.

5. Discussion

This study’s contribution lies in its ability not only to identify de-urbanization, but also to pioneer the identification of various pathways of de-urbanization and conduct a spatial exploration and a spatiotemporal analysis. A total of 727,375 pixels were detected to have undergone a transition from construction land to non-construction land, indicating that de-urbanization is a significant and noteworthy phenomenon. This finding provides substantial evidence for the effectiveness of China’s Ecological Civilization Construction and Stock Development strategies. Although some current studies on land use change have detected de-urbanization, researchers have not analyzed it [39,40].

Based on the findings from the correlation analysis between de-urbanization and distance from the city center, it is evident that de-urbanization is less prevalent in the core area, which challenges the traditional understanding of de-urbanization. Previous perspectives commonly held that de-urbanization, as a form of redevelopment, would predominantly occur in the core area due to the abundance of aging structures. The study conducted by Wu indicates that regreening tends to happen more frequently in the older, established parts of cities [20]. However, our study reveals that de-urbanization is relatively rare in the core area, instead clustering in areas located 10–30 km away from the city center. This finding aligns with the conclusions of Ni’s (2023) study, which identified that urban renewal areas are predominantly situated in the suburbs surrounding clusters of old urban areas [6]. Although their study focuses on urban renewal areas, distinct from de-urbanization in our study, both reflect similar significance and patterns. According to the Rent Gap Theory and the Bid Rent Theory, the land rent gap tends to increase as the distance from the urban core area decreases. This implies that there is a greater potential for redevelopment in the urban core area [41]. However, the transaction costs of redevelopment cannot be ignored. The core area of a city is typically characterized by the highest levels of urban land rent, accompanied by significantly dense building and population concentrations, which incur substantial transaction costs. High-level transaction costs, including large amounts of compensation for demolition and relocation, as well as a great deal of time spent on negotiation and coordination, would hinder the progress of redevelopment activities. By contrast, the area surrounding the core area, with high land rent gap levels and relatively low transaction costs, become optimal choices for redevelopment activities [42,43].

The aforementioned findings have significant policy implications for urban development. Since there are challenges associated with redevelopment in the core area, it is crucial for newly developing cities to prioritize effective, cautious urban planning, controlled construction density, and rational land use arrangements. By doing so, they can prevent potential difficulties in future redevelopment endeavors.

De-urbanization can reflect local governments’ considerations of the economic and ecological benefits of land use. The increased consideration of reverse changes in construction land during redevelopment indicates that local authorities place greater emphasis on high-quality development rather than purely economic returns.

This study has several limitations. Firstly, the merging of grassland and agricultural land into a single category labeled as “Low plants” obscures the distinct changes in these two important land types. This study identified that transformation into “Low plants” is the most significant type of de-urbanization in the GBA. However, the distribution of grassland and agricultural land within cities is inherently diverse, and their ecological and social impacts on urban areas vary significantly [44,45,46]. A more detailed classification would better illustrate the diversity of urban land reverse change pathways. Future research could consider separating these different pathways and analyzing the distribution characteristics and impacts of agricultural land and grassland individually.

Additionally, this study did not conduct a comprehensive analysis of driving mechanisms in a quantified way. De-urbanization results from the combined effects of multiple factors. Natural indicators, such as terrain and slope, as well as socio-economic indicators, like GDP and population, should be included in the analysis of driving forces. Furthermore, it is important to investigate the extent to which GDP influences de-urbanization and how urban renewal policies and pandemic-related disruptions affect redevelopment patterns. These topics warrant further in-depth exploration in future research.

6. Conclusions

Urbanization and de-urbanization are both typical processes of land use/cover change. While there has been extensive and in-depth research on urbanization, studies on de-urbanization are relatively scarce. This study investigated de-urbanization in the GBA from 2014 to 2023, utilizing multi-source remote sensing data and the CCDC method. It identified the types and timing of changes for each unit and conducted an in-depth analysis based on spatiotemporal pattern characteristics. The conclusions drawn from this study are as follows:

(1) During the period from 2014 to 2023, a total of 72.74 km² of de-urbanization was detected in the GBA. The primary form of de-urbanization observed in the study was the transformation of construction land into areas comprising “Low plants,” which includes both grassland and agricultural land. There are significant variations in de-urbanization between different cities and within individual cities. These variations are influenced by factors including the economic development level, land area, and government policies.

(2) Temporally, the annual changes in deurbanization exhibit significant variations. De-urbanization primarily occurred in 2016, 2019, and 2021, with a sharp decline in 2022. This is closely related to the formulation and implementation of urban renewal policies, as well as significant impacts from the pandemic.

(3) The de-urbanization process in the GBA exhibits spatial agglomeration, primarily characterized by low-level clustering. This indicates that de-urbanization in the GBA is predominantly driven by small-scale transformations, which is closely related to the recent promotion of micro-redevelopment by the Guangdong province. Large-scale comprehensive transformations have gradually been phased out and are only observed in a few specific areas. The conclusions of this study will provide guidance for the formulation of urban renewal and urban governance policies.

(4) Location factors play a crucial role in influencing de-urbanization. The proximity to the city center and the distance from main roads indeed influence the distribution and density of de-urbanization. Specifically, de-urbanization tends to concentrate within a radius of 10–30 km from the city center, particularly in close proximity to major roads.

(5) The development of de-urbanization is closely related to the stage of urbanization. The trend of de-urbanization follows a pattern that is consistent with the Northam curve. During the rapid urbanization phase, de-urbanization tends to increase. As urbanization reaches a mature stage, the growth of de-urbanization tends to decline.