Monitoring Urban Expansion (2000–2020) in Yangtze River Delta Using Time-Series Nighttime Light Data and MODIS NDVI

Abstract

The Yangtze River Delta Urban Agglomeration (YRDUA), which is located in the convergence zone of “The Belt and Road Initiative”, is one of the regions with the best urbanization foundations in China. Referring to the four five-year plans (China’s national economic plan), this study aimed to investigate the spatiotemporal patterns of urban expansion in the YRDUA from 2000 to 2020. To conduct a long-term analysis of urbanization, an extended time series (2000–2020) of a nighttime light (NTL) dataset was built from the multi-temporal Defense Meteorological Satellite Program’s Operational Linescan System (DMSP-OLS) data (2000–2013), and Suomi National Polar-Orbiting Partnership Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) data (2014–2020); data from these sources are crucial to understanding the urbanization processes in the region in order for more effective decision making to take place. The support vector machine (SVM) method was used to extract urban clusters from the extended time-series NTL data and MODIS NDVI products. The evolution of the urban expansion intensity was detected at city scales, and the inequality of urban growth was demonstrated using the Lorenz curve and Gini coefficient. Finally, a quantitative relationship between urban NTL intensity and socio-economic data was built to explore the main factors that control urban intensity. The results indicated that the urban extents extracted from time-series NTL data were consistent with those extracted from Landsat data, with an average overall accuracy (OA) of 89%. A relatively fast urbanization pace was observed during the 10th five-year plan (from 2000 to 2005), which then declined slightly in the 11th five-year plan (from 2006 to 2010). By the 12th and 13th five-year plan (from 2011 to 2020), urban clusters in all cities tended to grow steadily. Urban expansion has presented a radial pattern around the main cities, with sprawl inequality across cities. The results further revealed that the primary factors controlling NTL brightness were gross domestic product (GDP), total fixed asset investment, tertiary industry, gross industrial output, urban area, and urban permanent residents in city clusters, but the same driving factors had a different contribution order on the NTL intensity across cities. This study provides significant insight for further urbanization study to be conducted in the YRDUA region, which is crucial for sustainable urban development in the region.

1. Introduction

Urbanization is a complicated process relating to rural–urban migration, socio-economic development, and population growth [1,2,3]. Since the reform and opening up policy in China in 1978, the country has undergone rapid economic development and an accelerated urbanization process [4,5,6,7]. For instance, the Yangtze River Delta Urban Agglomeration (YRDUA), which is considered the sixth largest urban agglomeration in the world, has experienced the highest level of urbanization. Such rapid urban expansion has resulted in a large number of environmental and ecological problems [8,9,10,11,12]. Therefore, timely and accurate information on the dynamics of urban development is significant for urban planning and decision making [13].

As recorded in the China City Statistical Yearbook published by the National Bureau of Statistics, urban areas include residential land, public facilities land, industrial land, storage land, external transportation land, road and square land, municipal facilities land, and special land [14]. Therefore, an urban area is mainly determined based on the characteristics of land use and the scope of human activities, which are important factors representing the urbanization process. At present, remote sensing has proven to be an effective technology for monitoring urban area change owing to its characteristics of all-weather, all-position, multi-platform, and multi-time [15,16,17]. Particularly, medium- and high-spatial-resolution remote sensing images, such as Landsat TM/ETM+, SPOT HRV, and IKONOS, have been widely employed in order to investigate urban sprawl [18,19,20,21]. However, limited geographic coverage and atmospheric conditions make it difficult to monitor urban expansion in a large-scale region such as the YRDUA.

NTL images with coarse spatial resolution, such as those from the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP-OLS) and the Suomi National Polar-Orbiting Partnership Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) [22,23,24,25], provide a novel perspective for investigating urban sprawl. NTL data can not only extract impervious surfaces by integrating other remote sensing data to obtain land cover information but can also analyze human activities by its special information collection capability [26]. Therefore, urban area extraction based on NTL data has received much attention.

Coarse spatial resolution images (500–1000 m) from the moderate-resolution imaging spectroradiometer (MODIS), with wide geographical coverage and high revisit frequency, have the potential to detect urban dynamics in large-scale areas [27,28]. The products of the normalized difference vegetation index (NDVI) derived from MODIS can reflect the dynamics of vegetation information inside and outside cities. Additionally, the physical properties of low-vegetation land surfaces are usually similar to those of urban land surfaces. Considering the negative correlation between vegetation density and urban centrality, some studies have taken MODIS NDVI as auxiliary data for extracting urban extents over large areas [29,30,31].

As the joint region of the national strategies of the “Belt and Road Initiative” and “Yangtze River Economic Belt Initiative”, the YRDUA plays a vital role in the economic development and opening up of China [4]. According to the data of the National Bureau of Statistics in 2020, the YRDUA has accounted for about 3.7% of the country’s land area, 16.7% of the country’s resident population, 36.9% of the country’s total imports and exports, and 24.1% of the country’s gross national product [14].

With the development of urbanization in China, the YRDUA has experienced a rapid urbanization process over the past twenty years. While the “Belt and Road Initiative” was proposed in 2013, Chinese government departments further recognized the importance of maintaining a steady and coordinated development of urbanization. For instance, the 13th five-year plan (2016–2020) emphasized the integration of urban and rural development and the importance of striving toward closing the narrow urban–rural development gap. Thus, referring the four five-year plans (China’s national economic plan) from 2000 to 2020, monitoring urban dynamics in an accurate and timely manner, urban sprawl and its driving forces can be assessed, which is of significance to promote urban planning and decision making to provide supports for sustainable urban development in the region.

The objectives of this study were: (1) to extract the urban areas of the YRDUA from 2000 to 2020 using the extended time-series NTL data (integrating DMSP/OLS and NPP-VIIRS) and MODIS NDVI; (2) to quantitatively analyze the spatiotemporal patterns of urban expansion in the YRDUA; and (3) to explore the driving factors of urban expansion in the YRDUA.

The rest of the paper is organized as follows: In Section 2, a review of relevant studies on monitoring urban extensions based on NTL data is presented. In Section 3, we describe the study area, data, and data pre-processing. Section 4 is the Methodology section, which includes extracting urban areas and analyzing urban expansion patterns and its driving forces. Section 5 is the Results section, where we evaluate the accuracy of the extraction of urban areas based on NTL data, investigate the spatiotemporal patterns of urban expansion, and explore the relationship between NTL intensity and urbanization variables. Section 6 is a discussion of the results and limitations, followed by a conclusion in Section 7.

2. Literature Review

Since the National Oceanic and Atmospheric Administration’s National Geoscience Data Center (NOAA/NGDC) created a digital archive of DMSP/OLS (extending from 1992 to 2013) [32], many studies have since measured large-scale urban expansion using nighttime light (NTL) data. For example, Elvidge et al. [33] extracted global impervious surface areas from DMSP/OLS images between 2000 and 2001. Sutton [34] measured the urban sprawl of the United States using DMSP/OLS data and gridded population density data. Using DMSP/OLS data in 2000, 2005, and 2010, Xie and Weng [35] detected urban extents in China using an object-based urban thresholding method. Xu et al. [36] used time-series DMSP/OLS data to analyze the spatiotemporal characteristics of urban expansion in China from 1992 to 2009 on regional and national scales. Zou et al. [37] investigated urban expansion patterns in the middle reaches of the Yangtze River based on DMSP/OLS data from 1992 to 2011.

Notably, the global NPP-VIIRS NTL images that have been released by NOAA/NGDC since 2013 are superior to DMSP/OLS data concerning spatial resolution and light imaging detection limits. Since then, NPP-VIIRS data have become a better option for mapping recent urban extents at different scales. Shi et al. [24] extracted the urban areas of twelve cities in China from 2012 VIIRS data and made comparisons with the extractions from DMSP/OLS data; the result showed that the urban areas extracted from VIIRS had a better accuracy. Guo et al. [27] mapped the impervious surface area distribution in China using an integration of 2012 NPP-VIIRS and MODIS NDVI data. Sharma et al. [28] extracted global urban areas using the fusion of NPP-VIIRS NTL images and MODIS data from 2014. These studies indicate that it is feasible to perform urban extent extraction and expansion analysis using nighttime light data.

Previous studies have proposed several methods to extract urban areas from NTL images, such as the threshold method [18,30], graphical mutation method [35], and index construction method [31,38]. Among these methods, the threshold method is easy to operate, but it is susceptible to the light blooming effect, resulting in a larger area of extracted urban land. As for the graphical mutation method, the threshold value of the boundary requires a number of experiments to determine that the extraction thresholds of various city boundaries are different, making it more costly to extract macro-scale urban areas. The index construction method reduces the light blooming effect by introducing NDVI data, but the expression form of different indices impacts the extraction results, and the unstable background also influences the accuracy of the results.

The support vector machine (SVM) method is a nonparametric classification method based on statistical learning theory [39]. Compared with traditional classification methods, the SVM method can combine urban- and non-urban-area features from different data as samples for training and classification. Thus, it has been widely used in the classification of remote sensing images [38,40]. To reduce the light bloom effect and improve the efficiency of long-time-series large-scale urban extraction, a few studies have employed the SVM method to extract urban areas from NTL images. For example, Zhang et al. [41] used the SVM algorithm to extract China’s urban land from DMSP/OLS, NDVI, and LST data. The results have proven that the classification accuracy obtained via the SVM algorithm is better than that from the local threshold optimization method.

The rapid urbanization process in the Yangtze River Delta has attracted several researchers to investigate urban dynamics using the NTL data in the region. Du et al. [42] mapped the urban expansion in the Yangtze River Delta based on DMSP/OLS data from 1992 to 2003. Using DMSP/OLS data, Wang et al. [43] detected urban sprawl as the driving force for Yangtze River Delta urban agglomeration from 1992 to 2009. Lu et al. [44] analyzed the characteristics of urban expansion in the YRDUA based on DMSP/OLS data in 1993, 1997, 2002, 2007, and 2012. However, due to the time availability of DMSP/OLS images, such studies mainly focused on urban expansion before 2012. Quantitative analysis of urban sprawl remains relatively rare using both DMSP/OLS and NPP-VIIRS NTL images simultaneously because of the difference between the two kinds of NTL images.

Therefore, to extract urban areas of the YRDUA from 2000 to 2020, in this study, we built an extended time series (2000–2020) of nighttime light (NTL) dataset, which was produced via cross-sensor calibration from DMSP-OLS NTL data (2000–2013) and a composition of monthly NPP-VIIRS NTL data (2014–2020). The SVM method was used to classify urban and non-urban areas from the extended dataset. Based on the extracted urban areas, we quantitatively analyzed the spatiotemporal patterns of urban expansion and its driving force to understand the constraints of urbanization across cities, so as to provide insights for healthy and orderly urban development in this region.

3. Study Area and Datasets

3.1. Study Area

The YRDUA is located in the developed eastern region of China and is one of the six internationally recognized urban agglomerations (Figure 1). According to the city cluster development plan in the Yangtze River Delta region approved by the State Council in May 2016, the urban agglomeration consists of the megacity of Shanghai, nine cities in Jiangsu province, eight cities in Zhejiang province, and eight cities in Anhui province, with a total land area of 211.7 thousand square kilometers [45]. The YRDUA is in the middle and lower reaches of the Yangtze River Plain area, with a relatively gentle topography. The urban agglomeration is near the East China Sea and the Yellow Sea with numerous lakes and abundant water resources. Relying on superior geographical location and national policies, transportation within the YRDUA is convenient with developed transport facilities for sea, land, and air travel. The advantages of superior geographical location and rich resources have made the YRDUA one of the most economically developed and most significant urban growth regions in China [5,6,8]. Rapid urbanization in the Yangtze River Delta Urban Agglomeration has profoundly impacted China’s economy and ecological environment.

3.2. Datasets

Five types of remote sensing data were used in this study: (1) DMSP/OLS (Defense Meteorological Satellite Program/Operational Linescan System) NTL images from 2000 to 2013; (2) NPP-VIIRS (Suomi National Polar-Orbiting Partnership/Visible Infrared Imaging Radiometer Suite) monthly composited data from 2012 to 2020; (3) MODIS (moderate-resolution imaging spectroradiometer) NDVI monthly composited products from 2000 to 2020; (4) Landsat TM images from 2002, 2004, 2005, 2009, and 2010; and (5) Landsat 8 OLI images from 2013, 2014, and 2015. We used the DMSP/OLS and NPP-VIIRS NTL data and MODIS NDVI products to extract urban areas in the YRDUA from 2000 to 2020. Landsat TM and OLI data were used to assess the NTL-derived urban events, because the spatial resolution of the Landsat data (30 m) is much finer than that of DMSP/OLS data (1 km) and VIIRS data (500 m). Considering the atmospheric conditions and geographic coverage at the time of image acquisition, Landsat TM and OLI data in different years were used to verify the accuracy of urban areas extracted from NTL data.

Table 1. The description of datasets.

Datasets	Time	Description	Data Source
DMSP/OLS NTL	2000–2013	Yearly images, with 30 arc second spatial resolution	The Earth Observation Group, NOAA’s National Geophysical Data Center. http://ngdc.noaa.gov/eog/dmsp/downloadV4composites.html (accessed on 1 January 2021)
NPP-VIIRS DNB	2012–2020	Monthly average radiance composite images from January to December, with 15 arc second spatial resolution	The Earth Observation Group, NOAA National Geophysical Data Center. http://www.ngdc.noaa.gov/dmsp/data/viirs_fire/viirs_html/viirs_ntl.html (accessed on 1 January 2021)
MODIS NDVI	2000–2020	Monthly composite, with 500 m spatial resolution	International Scientific and Technical Data Mirror Site, Computer Network Information Center, Chinese Academy of Sciences. http://www.gscloud.cn (accessed on 1 January 2021)
Landsat TM	2002, 2004, 2005, 2009, 2010	All bands are used except thermal infrared band, for which spatial resolution is coarse	International Scientific and Technical Data Mirror Site, Computer Network Information Center, Chinese Academy of Sciences. http://www.gscloud.cn (accessed on 1 January 2021)
Landsat 8 OLI	2013, 2014, 2015	Multispectral bands whose spatial resolutions are 30 m and 15 m are used	International Scientific and Technical Data Mirror Site, Computer Network Information Center, Chinese Academy of Sciences. http://www.gscloud.cn (accessed on 1 January 2021)

presents a brief description of the data sets.

In addition, in order to explore the quantitative correlation between urban expansion and economic growth, we also used the socio-economic indicator data published in the China Statistical Yearbook [14]. Administrative boundary data were obtained from the National Geomatics Center of China as auxiliary data (available at http://www.ngcc.cn, accessed on 1 January 2021).

3.3. Data Pre-Processing

To eliminate noise and errors, the original DMSP/OLS data required inter-calibration, intra-annual composition, and inter-annual correction [5] to improve the continuity and comparability of the data. Because NPP-VIIRS NTL data derived from the NGDC website had not eliminated the light detections related to gas flares, biomass burning, aurorae, and background noise [46,47,48,49,50], which impeded the accuracy of extraction for urban areas, we carried out a series of pre-processing procedures to minimize these adverse effects using the method proposed by Shi et al. [51]. The correction method included three main steps: (1) The initial corrected NPP-VIIRS data were obtained by generating a mask of NPP-VIIRS data from the pixels whose digital number (DN) values were positive in 2012 DMSP/OLS data. (2) The outliers were corrected according to the highest DN value of the three most developed megacities (Beijing, Shanghai, and Guangzhou) in China. (3) DN values less than zero were set to zero.

Since the monthly MODIS NDVI compositions had been radiometrically corrected, the maximum value composite method [52] was applied to produce the annual NDVI images using the following formula:

$${\mathrm{NDVI}}_{\max }=\max \left\{{\mathrm{NDVI}}_1,{\mathrm{NDVI}}_2,\dots ,{\mathrm{NDVI}}_{12}\right\}$$

(1)

where NDVI_max is the maximum image of multi-temporal NDVI images, and NDVI₁, NDVI₂, …, NDVI₁₂ are NDVI images acquired from January to December, respectively.

After inter-calibration, intra-annual composition, and inter-annual correction, the corrected NTL images were reprojected to the Albers Conical Equal Area projection and resampled to 500 m, and they were further used to extract the urban extents for each city.

4. Methodology

The methodology in this paper included three main parts: (1) extracting urban areas from 2000 to 2020 in the YRDUA; (2) analyzing the spatiotemporal changes of urban expansion; and (3) exploring the driving factors of urbanization (Figure 2).

4.1. Extraction of Urban Areas

DMSP-OLS stable nighttime light data and NPP-VIIRS nighttime light data have been widely used to monitor urban sprawl. However, due to their differences in spatial resolution and spectral response, the two datasets cannot be directly used together for an analysis of the urbanization process in the long term. Several studies have proposed extending time-series NTL data by integrating DMSP-OLS data and NPP-VIIRS data [51,52,53]. An extended time series (2000–2020) of DMSP/OLS-like NTL data was built, referring to the cross-sensor calibration proposed by Li [50], which simulated DMSP/OLS composites from the NPP-VIIRS (2013–2020) composites in this study.

The SVM method was adopted to extract urban areas from the extended NTL data and annual NDVI images. As a statistical learning method, the SVM method can efficaciously deal with high-dimensional data with finite training samples [39] and has been successfully employed in the classification of remote sensing images [38,40,54,55,56,57]. SVM is a binary model, which separates positive and negative training samples by finding a hyperplane, and it generates the maximum geometric interval between them. The learning strategy of the SVM method is interval maximization, which can be converted into solving the following constrained quadratic programming problem (Equation (2)) [58]:

$$\begin{array}{l}\mathrm{Maximize} W\left(\alpha \right)={\displaystyle \sum _{i=1}^n{\alpha }_i-\frac{1}{2}{\displaystyle \sum _{i=1}^n{\displaystyle \sum _{j=1}^n{\alpha }_i{\alpha }_j{y}_i{y}_{j}K\left(x_i,x_j\right)}}}\\ \mathrm{Subject} \mathrm{to} {\displaystyle \sum _{i=1}^n{\alpha }_i}{y}_i=0 \mathrm{and} 0\le {\alpha }_i\le c\end{array}$$

(2)

where i = 1, 2, …, n; x_i is the training sample vector; y_i is the category label; K(u, v) is the kernel function; and c is the penalty coefficient.

SVM method classifiers include linear classifiers and nonlinear classifiers [59]. The linear classifier finds a hyperplane in the two-dimensional space to maximize the interval between two kinds of samples, while nonlinear classifiers map data to higher-dimensional space via kernel function to solve the problem of linear indivisibility in original space (Figure 3).

The extraction urban area procedures included four main steps: (1) normalizing the data; (2) selecting urban and non-urban samples; (3) training the sample data using an iterative process of SVM classification; and (4) extracting urban areas and post-classification processing.

For the first step, to speed up the network convergence and improve the classification accuracy, the digital number (DN) values for DMSP/OLS, NPP-VIIRS, and NDVI data were scaled to the range of 0 to 1.

For the second step, we selected two training sets as the urban and non-urban samples using the threshold method. The densi-graph method [42] was employed to obtain the thresholds from contour maps, which were generated from the raster images of DMSP/OLS, NPP-VIIRS, and NDVI data [60]. Urban training samples were taken from the pixels with an NTL digital number (DN) value greater than the threshold of NTL data and an NDVI value smaller than the threshold of NDVI data, whereas the non-urban training samples were the opposite. Notably, the impact of water for which the NDVI value was smaller than zero was removed from the urban training samples.

For the third step, we built the SVM classification model by selecting the radial basis function (RBF) as the kernel function and used a grid search method to select the optimal values with five times cross-validation. Referring to the terminate condition in the iterative process for SVM training and classification in Ma et al. [57], if the current result shows an urban area that differs by no more than 8% from the previous one, the recursive invocation will cease. In addition, considering the continuous outward growth of urban areas in the region, we added a post-processing condition, in which urban clusters detected in previous years should be retained in subsequent years to ensure the continuity of urban areas extracted from different years.

Considering the heterogeneity of urban development across cities for different times, SVM training and classification were performed for each city rather than the whole region in this study.

4.2. Accuracy Assessment of Urban Areas

We assessed the accuracy of the urban areas extracted from NTL images by comparing them with those extracted from Landsat images. Specifically, we evaluated these extractions in terms of the similarity of the spatial extent and spatial shape, respectively.

The overall accuracy (OA), balanced accuracy (BA) [60], and Kappa coefficient were selected to serve as evaluating indicators for an assessment of the accuracy of the spatial extent.

Shape indictors such as the perimeter–area ratio (PARA), shape index (SHAPE), fractal dimension index (FRAC), related circumscribing circle (CIRCLE), and contiguity index (CONTIG) were used to estimate the spatial similarity of urban patches extracted from NTL data and Landsat images, respectively. A detailed description of these indices [61,62] is listed in

Table 2. The description of patch-shape indices.

Patch-Shape Index	Formula	Description
Perimeter–area ratio (PARA)	$\mathrm{PARA}=\frac{p_{ij}}{a_{ij}}$	pij and aij are perimeter and area of patch ij, respectively
Shape index (SHAPE)	$\mathrm{SHAPE}=\frac{0.25p_{ij}}{a_{ij}}$	pij and aij are perimeter and area of patch ij, respectively
Fractal dimension index (FRAC)	$\mathrm{FRAC}=\frac{2\ln \left(0.25p_{ij}\right)}{\ln a_{ij}}$	pij and aij are perimeter and area of patch ij, respectively
Related circumscribing circle (CIRCLE)	$\mathrm{CIRCLE}=1-\frac{a_{ij}}{a_{ij}^s}$	$a_{ij}$ and $a_{ij}^s$ are area of patch ij and its smallest circumscribed circle
Contiguity index (CONTIG)	$\mathrm{CONTIG}=\frac{\left(\frac{{\displaystyle \sum _{r=1}^z{c}_{ijr}}}{a_{ij}^{*}}\right)-1}{v-1}$	cijr is the contiguity value of pixel r in patch ij; v is the sum of the value in 3 × 3 cell template; and $a_{ij}^{*}$ is the total number of pixels in patch ij

Note: The formulas of the indicators are from the Fragstats v4.2.1 software.

.

4.3. Spatiotemporal Analysis of Urban Expansion

4.3.1. Urban Expansion Intensity

The urban expansion intensity (UEI) could reflect the urbanization levels in a region at different times [63,64]. To quantify the urbanization intensity for each city of the YRDUA in different periods, UEI was calculated every five years, responding to the 10th five-year plan (2000–2005), 11th five-year plan (2006–2010), 12th five-year plan (2011–2015) and 13th five-year plan (2016–2020), respectively. The formula used was as follows [63,64]:

$$\mathrm{UEI}=\frac{U_b-U_a}{U_a}\times \frac{1}{T}\times 100$$

(3)

where UEI is the urban expansion intensity; U_a and U_b are urban areas at the times of year a and year b, respectively; and T is the time span between a and b, i.e., five.

4.3.2. Inequality of Urban Growth

Identifying the unequal distribution of urban growth is crucial to optimize the spatial layout of cities. The Lorenz curve proposed by M. Lorenz [65] in 1905 was used to quantify unevenness in the distribution of urban expansion in the YRDUA in this study, which uses frequency accumulations to plot the degree of inequality (centralized or decentralized). The Lorenz curve is an outward convex curve, and when it is at an angle of 45° with the abscissa, it is called the line of equality. When the Lorenz curve is applied to evaluate the unevenness in the distribution of urban expansion in the YRDUA, the location entropy of the urban areas of 26 cities in the YRDUA in 2000, 2005, 2010, 2015, and 2020 was firstly calculated. The location entropy was calculated as shown in Equation (4) [66]:

$$Q_i=\left(U_i/U\right)/\left(L_i/L\right)$$

(4)

where Q_i is the location entropy of the i-th city; U_i is the urban area of the i-th city; U is the urban area of the whole YRDUA; L_i is the land area of the i-th city; and L is the total land area of the YRDUA.

After the calculation was completed, the location entropy was sorted from low to high. Additionally, the percentage of each city’s land area to the YRDUA’s total land area and the percentage of each city’s urban area to the YRDUA’s total urban area was calculated. Finally, the Lorenz curve was drawn with the cumulative percentage of land area as the horizontal coordinate and the cumulative percentage of urban area as the vertical coordinate (Figure 3). The distance between the curve and the line of equality is the difference between the actual distribution of urban area and its uniform distribution in the whole region. The closer the curve is to the line of equality, the more evenly distributed urban areas are in the whole region. On the contrary, the farther away from the line of equality the curve is, the more unbalanced the distribution of urban areas in the whole region.

However, the Lorenz curve only presented a rough estimation of the imbalance degree. Thus, we used the Gini coefficient [67,68] as another indicator to further quantify the unevenness of the distribution of urban areas in the study area. This can be expressed as the ratio A/(A + B) in Figure 4 (A represents the area of the gray plot and B represents the area of the blue plot). Generally, a Gini coefficient value larger than 0.4 indicates a dramatic difference in the distribution of urban areas [69]. A change in the value of the Gini coefficient can exhibit a trend of unevenness of urban growth.

4.4. Quantitative Relationship between NTL Intensity and Urbanization Factors

Time-series NTL intensity can indirectly reflect urbanization level as a recorder of anthropogenic luminosity. Diverse statistical models have been put forward to explore the relationship between NTL brightness and urbanization indicators, such as population, economic activity, and land-use change [37]. The contribution of various urbanization factors to the variations in NTL brightness has attracted much attention. In this study, we attempted to investigate the quantitative relationship between potential driving factors of urbanization and nighttime light brightness in Shanghai (municipality) and three capital cities, Nanjing, Hangzhou, and Hefei.

Referring to the selection of driving factors of urban expansion in the coastal region [68,70], we selected six urbanization indicators (GDP, permanent urban residents, gross industrial output, total fixed asset investment, tertiary industry values, and urban areas) to carry out partial correlation analysis with NTL brightness.

Firstly, the Z-score method (Equation (5)) [71] was used to normalize the urbanization factors to keep them on the same measurement scale:

$$Z=\frac{X_{\mathrm{i}}-\overline{X}}{S}$$

(5)

where Z is the normalized variable, X_i is the i-th variable, $\overline{X}$ is the average of the variable, and S is the standard deviation of the variable.

To simplify the description, we used ZL, ZD, ZP, ZI, ZA, ZT, and ZU to denote the normalized variables corresponding to total NTL intensity, GDP, permanent urban residents, gross industrial output, total fixed asset investment, total tertiary industry values, and urban areas, respectively.

Then, considering the correlations between the urbanization variables, we employed the principal component analysis method to eliminate the collinearity between variables [72]. After the selected driving factors were normalized, principal components were generated.

The principal components were calculated as [37]:

$$F1=w_{1}ZD+w_{2}ZP+w_{3}ZI+w_{4}ZA+w_{5}ZT+w_{6}ZU$$

(6)

where F1 represents principal components; ZD, ZP, ZI, ZA, ZT, and ZU are normalized variables corresponding to GDP, permanent urban residents, gross industrial output, total fixed asset investment, total tertiary industry values, and urban areas, respectively; and w_i represents the corresponding elastic coefficients.

Finally, a linear relationship between the principal components and NTL intensity was established using the least square method to analyze the influence of various urbanization factors on the evolution of NTL intensity.

5. Results

5.1. Urban Sprawl and Accuracy Assessment

5.1.1. Distribution Pattern of Urban Sprawl

Figure 5 presents the NTL-derived urban areas of the YRDUA in 2000, 2005, 2010, 2015, and 2020. As can be noticed, the YRDUA experienced a significant urban expansion from 2000 to 2020, but the characteristics of urban expansion were diverse across cities in different periods. The urban expansion was “Z-shape” along the main cities of Hefei, Nanjing, Suzhou, Shanghai, Hangzhou, and Ningbo. “Z-shape” refers to the spatial distribution characteristics of urban areas. This distribution of urban space contributes to a spatial association and radiating effect among cities, especially the radiating effect of Shanghai, which is located at the inflection point of the region. Cities in the “Z-shape” are the main force of urbanization development, including Shanghai, Hangzhou, Nanjing, Hefei, and other regional core cities, together forming a relatively balanced network structure. This spatial distribution of urban expansion helps to strengthen the spatial connection between cities and has a radiative effect on core cities.

5.1.2. Accuracy Assessment

Figure 6 shows the difference between the urban areas extracted from DMSP/OLS and those extracted from DMSP/OLS-like data in 2012 and 2013 (the overlapping period of data collected by the two sensors). The results revealed that the two extracted urban areas were very similar both in terms of the spatial extent and spatial shape in this study. The findings also showed that the extended time-series DMSP-like NTL data with a better consistency was appropriate for further analyzing the urban spatiotemporal characteristics.

The relative errors between DMSP/OLS NTL intensities and DMSP/OLS-like NTL intensity for urban areas in 2012 or 2013 were all found to be less than 1%, indicating that the method of simulating DMSP/OLS NTL intensities is feasible (

Table 3. The category of urban expansion intensity.

NTL Intensity of DMSP/OLS in Urban Area	Shanghai	Nanjing	Hangzhou	Hefei
DMSP/OLS NTL intensity in 2012	335,185.66	171,025.44	121,285.41	97,901.59
DMSP/OLS-like NTL intensity in 2012	335,808.47	171,243.8	120,812.34	98,555.59
Relative error	0.19%	0.13%	−0.39%	0.67%
DMSP/OLS intensity in 2013	377,198.53	181,345.86	123,065.57	97,470.05
DMSP/OLS-like NTL intensity in 2013	379,605.56	182,268.34	124,180.56	97,311.727
Relative error	0.64%	0.51%	0.91%	−0.16%

).

Figure 7 and Figure 8 present the accuracy assessment results with the overall accuracy (OA) of 89%. They reveal that the urban areas extracted from NTL data were reasonably consistent with those from Landsat data.

Table 4. Comparison of spatial shape indicators for urban patches extracted from DMSP/OLS data and those extracted from Landsat TM.

City	Shape Index		Fractal Dimension Index		Perimeter–Area Ratio		Related Circumscribing Circle		Contiguity Index
	DMSP/OLS	TM	DMSP/OLS	TM	DMSP/OLS	TM	DMSP/OLS	TM	DMSP/OLS	TM
Nanjing	1.430	1.765	1.043	1.060	38.611	40.367	0.556	0.612	0.463	0.441
Suzhou	1.622	2.571	1.059	1.109	19.740	20.869	0.534	0.527	0.724	0.725
Wuxi	1.437	1.634	1.041	1.054	42.276	45.165	0.538	0.560	0.415	0.385
Changzhou	1.437	1.702	1.047	1.057	25.290	28.202	0.557	0.528	0.647	0.612
Nantong	1.346	1.845	1.042	1.072	24.855	24.796	0.508	0.559	0.653	0.663
Ma’anshan	1.514	1.730	1.056	1.064	32.447	36.337	0.602	0.570	0.546	0.499
Huzhou	1.287	1.580	1.035	1.055	34.266	33.072	0.500	0.552	0.525	0.548
Jinhua	1.357	1.862	1.044	1.073	29.656	28.303	0.497	0.586	0.624	0.610
Tongling	1.734	1.917	1.066	1.075	24.941	26.642	0.621	0.561	0.654	0.648
Chizhou	1.131	1.175	1.022	1.026	38.811	40.262	0.509	0.451	0.447	0.445

and

Table 5. Comparison of spatial shape indicators for urban patches extracted from the extended DMSP-like nighttime light (NTL) data and those extracted from Landsat OLI.

City	Shape Index		Fractal Dimension Index		Perimeter–Area Ratio		Related Circumscribing Circle		Contiguity Index
	DMSP-Like	OLI	DMSP-Like	OLI	DMSP-Like	OLI	DMSP-Like	OLI	DMSP-Like	OLI
Shanghai	1.400	1.784	1.042	1.057	41.610	43.792	0.546	0.583	0.425	0.399
Nanjing	1.455	1.892	1.045	1.060	37.074	37.010	0.557	0.567	0.487	0.486
Ningbo	1.314	1.548	1.032	1.047	34.257	36.753	0.517	0.461	0.526	0.512
Changzhou	1.506	1.722	1.045	1.058	32.715	28.926	0.567	0.452	0.545	0.611
Ma’anshan	1.263	1.498	1.031	1.055	37.784	38.714	0.495	0.589	0.473	0.462
Zhenjiang	1.573	1.690	1.055	1.049	32.833	37.296	0.567	0.352	0.540	0.517
Jinhua	1.628	2.075	1.058	1.085	24.448	25.981	0.517	0.587	0.661	0.652
Huzhou	1.359	1.578	1.042	1.055	24.843	20.653	0.509	0.496	0.653	0.712
Xuancheng	1.370	1.496	1.053	1.047	31.891	34.193	0.566	0.573	0.557	0.526
Chizhou	1.207	1.313	1.034	1.041	34.605	34.271	0.530	0.534	0.521	0.531

show the results of spatial shape indicators for the urban patches extracted from NTL images through comparison with those classified from Landsat data. The small average error indicates that the spatial shapes of the urban patches extracted from the two data are similar.

The relative error of the shape index is larger compared with those of other indices. A possible reason for this is that some small urban patches can be identified from Landsat data (finer spatial resolution), but not from NTL data. The comparison results of all shape indices revealed that the spatial shapes of urban patches extracted from NTL data were reasonably consistent with those classified from Landsat data.

5.2. Spatiotemporal Variations in Urban Expansion

5.2.1. Dynamic of Urban Expansion Intensity

Figure 9 exhibits the urban expansion intensity of the YRDUA during the 10th five-year plan (2001–2005), 11th five-year plan (2006–2010), 12th five-year plan (2011–2015), and 13th five-year plan (2016–2020). The results revealed that the YRDUA experienced a rapid urbanization process from 2000 to 2005, but there was a decline during 2006–2010. The incidences of urban sprawl were indirectly related to domestic and international events. China’s accession to the WTO in 2001 had a notable impact on industrial structure adjustment, economic development, and urban expansion, particularly in the YRDUA. However, the global financial crisis in 2008 resulted in a deceleration of urbanization. Nevertheless, accelerated expansion still can be found in Chizhou and Ningbo owing to the development of tourism in Chizhou and the harbor construction of Ningbo, which was supported by the work report of Chizhou Municipal People’s Government [73] and Ningbo Municipal Development and Reform Commission [74]. In 2011–2020, the urban expansion of the cities in the YRDUA showed a gradual downward trend, which implicates a smooth urbanization process in this period.

5.2.2. Unequal Distribution of Urban Growth

Figure 10a displays the Lorenz curves of urban areas for the YRDUA in 2000, 2005, 2010, 2015, and 2020. To show the changes in the distribution of urban areas during this 20-year period more visually, we added Figure 10b, which displays the Lorenz curves in 2000 and 2020 only. The distance between the equality line and the Lorenz curve represents the unevenness of the urban area distribution from 2000 to 2020. It can be observed that Lorenz curves are far away from the perfect equality line over the period, as the distribution of urban areas in the region has been relatively uneven (Figure 10a). However, the distances between them are gradually closer from 2000 to 2020, which reveals that this uneven distribution was moving in a positive direction (Figure 10b).

However, the Lorenz curve only presents a rough estimation of the imbalance degree. Thus, we calculated the Gini coefficient from the plotted Lorenz curve [64,65] as another indicator to further quantify the unevenness of the distribution of urban areas in the study area. The Gini coefficient decreased from 0.532 in 2000 to 0.473 in 2020 (Figure 10a). These results indicated that the distribution of urban areas in this region was not uniform and the proportion of urban areas among cities varied greatly. This may be attributed to the inherent advantages of large-scale cities over small-scale cities in terms of economy and industry, resulting in different levels of urbanization across cities. However, such disparities have gradually been reduced with the steady and coordinated development of urbanization in the YRDUA.

5.3. Relationship between NTL Intensity and Urbanization Variables

In this study, the value of NTL intensity in a city was defined as the total of DN values within the corresponding NTL-extracted urban area. Figure 11 shows the time-series NTL intensity curves within the city clusters of Shanghai (municipality) and Nanjing, Hangzhou, and Hefei (capital cities) from 2000 to 2020, respectively. As indicated in Figure 11, due to the different urbanization levels, the NTL intensities exhibited a positive but distinct trend. In general, the NTL intensity kept a steadily increasing trend from 2000 to 2020.

The partial correlation coefficients for all variables for the four cities are shown in

Table 6. Correlation coefficients between urbanization variables in Shanghai, Nanjing, Hangzhou, and Hefei.

City	Variables	ZL	ZD	ZP	ZI	ZE	ZT	ZU
Shanghai	ZL	1
	ZD	0.980	1
	ZP	0.888	0.865	1
	ZI	0.950	0.962	0.874	1
	ZE	0.972	0.964	0.880	0.972	1
	ZT	0.968	0.994	0.836	0.927	0.941	1
	ZU	0.974	0.955	0.891	0.981	0.980	0.921	1
Nanjing	ZL	1
	ZD	0.872	1
	ZP	0.973	0.840	1
	ZI	0.904	0.986	0.878	1
	ZE	0.882	0.995	0.855	0.994	1
	ZT	0.858	0.999	0.823	0.980	0.991	1
	ZU	0.999	0.883	0.973	0.914	0.892	0.869	1
Hangzhou	ZL	1
	ZD	0.971	1
	ZP	0.871	0.785	1
	ZI	0.968	0.949	0.783	1
	ZE	0.951	0.993	0.771	0.908	1
	ZT	0.922	0.968	0.728	0.874	0.976	1
	ZU	0.999	0.972	0.870	0.967	0.953	0.924	1
Hefei	ZL	1
	ZD	0.972	1
	ZP	0.977	0.942	1
	ZI	0.961	0.997	0.919	1
	ZE	0.979	0.990	0.956	0.982	1
	ZT	0.975	0.999	0.950	0.994	0.990	1
	ZU	0.997	0.966	0.973	0.954	0.974	0.966	1

Note: ZL, ZD, ZP, ZI, ZE, ZT, and ZU are the normalized variables corresponding to total NTL intensity, GDP, urban permanent residents, gross industrial output, total fixed asset investment, tertiary industry, and urban areas extracted from NTL data, respectively.

. The correlation coefficients between NTL intensity and the driving variables are all larger than 0.75 (Table 6), indicating that the driving factors selected in this study can be used as explanatory variables and also implying high correlations between the independent variables. Thus, the principal component analysis method is used to eliminate the collinearity between variables.

For Shanghai, we calculated the principal components as:

$$F{1}_{SH}=0.4148ZD+0.3858ZP+0.4131ZI+0.4148ZA+0.4064ZT+0.4139ZU$$

(7)

For Nanjing, we calculated the principal components as:

$$F{1}_{NJ}=0.4144ZD+0.3886ZP+0.4177ZI+0.4160ZA+0.4114ZT+0.4009ZU$$

(8)

For Hangzhou, we calculated the principal components as:

$$F{1}_{HZ}=0.4227ZD+0.3655ZP+0.4086ZI+0.4180ZA+0.4082ZT+0.4236ZU$$

(9)

For Hefei, we calculated the principal components as:

$$F{1}_{HF}=0.4113ZD+0.4006ZP+0.4080ZI+0.4109ZA+0.4118ZT+0.4068ZU$$

(10)

Quantitative relationships between the principal component F1 of the driving factors and the normalized NTL intensity (ZL) were obtained using the least square method (Figure 12).

Taking Shanghai as an example, the linear relationship between ZL and the principal component F1 is established as follows, and the R² of the model is 0.97:

$$Z{L}_{SH}=0.4146F{1}_{SH}$$

(11)

The quantitative relationship between the driving factors and ZL was further obtained:

$$Z{L}_{SH}=0.1720ZD+0.1600ZP+0.1713ZI+0.1720ZA+0.1685ZT+0.1716ZU$$

(12)

Similarly, the quantitative relationship between ZL and the driving factors for Nanjing (Equation (13)), Hangzhou (Equation (14)), and Hefei (Equation (15)) were obtained by using principal component analysis and linear regression analysis:

$$Z{L}_{NJ}=0.1648ZD+0.1545ZP+0.1661ZI+0.1654ZA+0.1636ZT+0.1594ZU$$

(13)

$$Z{L}_{HZ}=0.1789ZD+0.1547ZP+0.1729ZI+0.1769ZA+0.1728ZT+0.1793ZU$$

(14)

$$Z{L}_{HF}=0.1682ZD+0.1638ZP+0.1669ZI+0.1681ZA+0.1684ZT+0.1664ZU$$

(15)

These quantitative model results reveal the influence order of the six urbanization factors on the brightness of NTL. GDP and total fixed asset investment have equally greater influences on NTL brightness compared to other factors in Shanghai. However, the first influencing factor is total fixed asset investment in Nanjing, followed by GDP and tertiary industry, and in Hangzhou, urban area is the most important influencing factor, followed by GDP and gross industrial output. In Hefei, tertiary industry, GDP, and gross industrial output are the primary driving factors. Obviously, the same factors have a different contribution order in terms of the NTL intensity across cities. Various socio-economic driving factors jointly promote the steady development of urbanization in the YRDUA.

6. Discussion

NTL data can detect artificial lights, which provides a more specific perspective of urban dynamics. Previous studies have demonstrated that DMSP/OLS and NPP-VIIRS data had the potential to extract urban areas on large scales, respectively. However, in previous studies of the YRDUA, quantitative analysis of urban sprawl using both DMSP/OLS and NPP-VIIRS NTL images simultaneously remains relatively rare because of the difference between the two kinds of NTL images. In this study, we built an extended time series (2000–2020) of DMSP/OLS-like NTL data from DMSP/OLS (2000–2013) and NPP-VIIRS (2014–2020) together for a long-term analysis of urbanization. The findings indicate that the extended time-series DMSP-like NTL data exhibit a better consistency, which is appropriate for analyzing the urban spatiotemporal characteristics.

The SVM method was performed to extract urban extents of the YRDUA in 2000–2020 by employing the extended time-series NTL data and MODIS NDVI data. To access the urban clusters extracted from the extended time-series NTL data, we compared them with the urban areas delineated from finer-resolution Landsat images. The results show that the urban extents that are NTL-derived are consistent with those classified from Landsat images with an average OA of 89% and an average BA of 75%. Through the calculation of shape indicators such as perimeter-area ratio and shape index, we discovered a strong resemblance between the spatial shape of urban patches identified from NTL data and those derived from Landsat data. These findings prove that it is feasible to extract urban extents from the extended time-series DMSP-like NTL data. After parameter optimization using the grid search, the SVM model presented a good ability to classify urban patches from different levels of cities.

However, despite the merits of this study, there were still some limitations and uncertainties in the process of urban area extraction. On one hand, some pseudo-urban pixels in the suburban regions were misclassified into urban areas due to the characteristic of high NTL DN value and low vegetation density. On the other hand, NPP-VIIRS data were converted into DMSP-OLS-like data on the basis of the cross-sensor correction method, which traded off the quality of NPP-VIIRS nighttime light data. However, it may contain the error of the assumption that the urban areas detected in previous years remained in the following years in the post-processing of the data. In addition, considering the urban expansion discrepancy across cities, it is necessary to investigate the genericity and applicability of the whole method for urban expansion analysis for different cities combining more multi-source data such as high-resolution remote sensing imagery in future studies.

From the spatiotemporal characteristics of urban sprawl viewpoint, our results reveal the spatiotemporal heterogeneity of urban sprawl. Urban expansion in the YRDUA presented roughly “Z-shape” along Hefei, Nanjing, Suzhou, Shanghai, Hangzhou, and Ningbo during 2000–2020, which implies the associated spatial radiating effect among cities, especially in terms of the core radial effect of Shanghai as a megacity. Such a pattern was not only conducive to Shanghai playing the core role but was also beneficial to the radiating and driving roles of sub-central cities regarding the development of surrounding small towns. We analyzed the urban expansion intensity and the unequal distribution of urban growth quantitatively. The results show that the urban expansion intensity in the YRDUA was relatively larger during the 10th five-year plan (2000–2005) and presented a declining trend in the 11th five-year plan (2006–2010). A possible reason for this was that China became a member of WTO in 2001, which had a significant effect on industrial structure adjustment, economic development, and urbanization. Especially in the YRDUA, increasing import and export trade and foreign investment accelerated the urbanization process until 2008, when the financial crisis led to a slowdown of foreign trade exports and the manufacturing industry, negatively impacting the urban growth of the YRDUA. During the 12th and 13th five-year plans (2011–2020), most cities experienced slow but steady urban expansion. Meanwhile, the change in the Lorenz curve and Gini coefficient indicates that the inequality of urban growth had gradually decreased within the YRDUA across the fourth five-year plan due to the policy of urbanization coordinated development in the national economic plans in China.

Furthermore, we built a quantitative relationship between urban NTL intensity and urbanization driving factors based on socio-economic data. Our results reveal that GDP, total fixed asset investment, tertiary industry, urban area, and gross industrial output were important contribution factors toward the urban NTL brightness in the different levels of cities, while the contribution order of the urbanization factors differed across cities, although the socio-economic data may have some errors.

In summary, referring the four five-year plans from 2000 to 2020 in China’s national economic plan, our findings prove that the spatiotemporal pattern of urban expansion fully reflects the national policies for urbanization development in the YRDUA. A timely and explicit understanding of urban sprawl and its driving forces is important for urban planning and more effective decision-making analysis. The study provides important insights into future urbanization processes to be conducted for sustainable urban development in the YRDUA region.

7. Conclusions

In this study, we aimed to detect the spatiotemporal patterns of urban expansion and the driving factors for the YRDUA using time-series NTL images and MODIS NDVI data during 2000–2020. For long-term urbanization analysis, an extended time series (2000–2020) of DMSP/OLS-like NTL data was built using cross-sensor calibration from DMSP-OLS NTL data (2000–2013) and a composition of monthly NPP-VIIRS NTL data (2014–2020). Urban clusters were extracted using the SVM method, with an overall accuracy (OA) of 89%. The results revealed that the YRDUA underwent a rapid urbanization process during 2000–2020. Urban sprawl exhibited a “Z-shape” distribution along main cities, such as Shanghai, Nanjing, Hangzhou, Hefei, and Ningbo. Referring to the four five-year plans of the national economic development of China, urban expansion was relatively quick during the 10th five-year plan (2000–2005), but it declined in the 11th five-year plan (2006–2010). By the 12th and 13th five-year plans (2011–2020), urban areas in most cities tended to grow slowly. The spatiotemporal heterogeneity of urban expansion existed in cities of different levels. Meanwhile, the results indicated that the unevenness of urban growth had gradually decreased across temporal stages. Overall, the urban clusters exhibited a steady and radial expansion model. National economic development and plan support for urbanization are the main reasons for urban expansion. This study explored the relationship between NTL intensity and six selected urbanization variables for the municipality of Shanghai and three capital cities. The results indicated that NTL brightness is affected by socio-economic factors, GDP, total fixed asset investment, tertiary industry, urban areas, and gross industrial output, which were the primary controlling factors.

In future studies, the following issues should be considered: (1) the incorporation of more multi-source data to investigate urban expansion patterns (such as high-resolution remote sensing imagery and Point of Interest data) and (2) adaptability of more machine learning methods in extracting data for urban areas across cities with urban expansion discrepancies.