Prediction of Multi-Scale Socioeconomic Parameters from Long-Term Nighttime Lights Satellite Data Using Decision Tree Regression: A Case Study of Chongqing, China

Abstract

The Defense Meteorological Satellite Program Operational Linescan System (DMSP/OLS) and the Suomi National Polar-Orbiting Partnership satellite’s Visible Infrared Imaging Radiometer Suite (NPP/VIIRS) nighttime light (NTL) data provide an adequate proxy for reflecting human and economic activities. In this paper, we first proposed a novel data processing framework to modify the sensor variation and fit the calibrated DMSP/OLS data and NPP/VIIRS data into one unique long-term, sequential, time-series nighttime-lights data at an accuracy higher than 0.950. Both the supersaturation and digital value range have been optimized through a machine learning based process. The calibrated NTL data were regressed against six socioeconomic factors at multi-scales using decision tree regression (DTR) analysis. For a fast-developing city in China—Chongqing, the DTR provides a reliable regression model over 0.8 (R²), as well explains the variation of factor importance. With the multi-scaled analysis, we matched the long-term time-series NTL indices with appropriate study scale to find out that the city and sub-city region are best studied using NTL mean and stander derivation, while NTL sum and standard deviation could be better applied the scale of suburban districts. The significant factor number and importance value also vary with the scale of analysis. More significant factors are related to NTL at a smaller scale. With such information, we can understand how the city develops at different levels through NTL changes and which factors are the most significant in these development processes at a particular scale. The development of an entire city could be comprehensively explained and insightful information can be produced for urban planners to make more accurate development plans in future.

1. Introduction

Remotely sensed nighttime light (NTL) image data can provide a unique perspective for observing human economic activities on a global scale and over a long period. Empirical studies have applied such data to research the urbanization process and its interaction with the environment, estimate population and GDP growth, measure energy consumption, and assess the impacts of urban expansion [1,2,3,4,5,6,7]. Hence, understanding the details of NTL data in a proper and correct format and applying them to downstream tasks is a critical challenge in current related studies [8,9,10,11,12,13,14,15]. The two most commonly used NTL datasets are from OLS and VIIRS images, both created by the Earth Observation Group (EOP) at NOAA. The Operational Linescan System (OLS) is an oscillating scan radiometer with low-light visible and thermal infrared imaging capabilities that first flew on DMSP satellites in 1976. The sensor has high-performance photoelectric amplification capabilities to detect and record city lights, natural fires, and even traffic flow in a city [4,5,6,7,8]. This series of night lights, as a characterization of human activities, became reliable data for human activity monitoring research when the DMSP/OLS ceased to acquire images. The creator switched to the new generation VIIRS (Visible Infrared Imaging Radiometer Suite) images from the Suomi National Polar Partnership satellite (SNPP), and the new sensor presents a significant improvement over the DMSP/OLS sensor, in both data availability and spatial resolution, from earth degree to 1 km [16]. In addition, the new VIIRS NTL data are radiometrically calibrated to prevent lower light sensitivity issues and saturation in urban areas or over-glow [17,18]. Therefore, the VIIRS images can provide a more accurate NTL data source for economic modeling.

To analyze and explore the economic and societal development dynamics at different scales, a global long-term, uninterrupted and continuous dataset of a fine temporal resolution is needed [19,20,21,22]. The NTL data seems to be a good solution to this problem, which can receive images from every place worldwide and over a long period from 1992 to the present. However, due to the sensors’ variation and the changing data format, the NTL data from OLS and VIIRS must be calibrated before being applicable to specific studies [22]. To address this issue, many data calibration methods have been developed in previous studies [21,22]. For example, based on the long-term OLS series dataset integrated with the VIIRS data from 2005 to 2019, Li et al. (2017) measured the spatio-temporal evolution characteristics of carbon emissions at or above the municipal level in the Beijing-Tianjin-Hebei region, providing a reference for the formulation of low-carbon development policy [23]. Based on long-term sequence lighting data from 1992 to 2018 for the Chinese mainland, Yong et al. (2022) analyzed the influencing factors of spatio-temporal changes and fundamental population changes in 38 important cities during this time [24]. Most of these researchers only focused on the single NTL type of sensor variation problem but ignored the inter-calibration process between DMSP/OLS data and SNPP/VIIRS data since both detecting sensors and the data acquisition and processing algorithms are different [25]. Li et al. promoted two benchmark pieces of research to light up this ignorance in 2017 [23] and introduced a model that used a power function to fit the non-linear relationship between VIIRS radiance and DMSP/OLS digital number and used a Gaussian low-pass filter for smoothing the integrated data. In 2020, Ma et al. [26] transformed the VIIRS and DMSP-OLS pixels to scatter plots and changed the power function to a Biphasic Dose Response model to build the calibration relationship. Chen et al. (2021) built an extended time-series (2000–2018) NPP-VIIRS-like NTL dataset through a new cross-sensor calibration from DMSP-OLS NTL data (2000–2012) and a composition of monthly NPP-VIIRS NTL data (2013–2018). The calibrated DMSP–OLS NTL data have a better temporal consistency and record the digital number (DN) values with a range from 0 to 63 [27]. Yong et al. (2022) combined the PSO-BP algorithm to investigate the relationship between the DMSP-OLS data and NPP-VIIRS data, and then integrated the two datasets at the pixel-level, and established a consistent time-series NTL dataset during the period of 2000–2019 [28]. Li et al. (2019) used the corrected DMSP/OLS image data to reclassify VIIRS/DNB, and then used the overlapping data of the two in time and space to linearly fit VIIRS/DNB image data [29]. Jiang et al. (2021) processed the DMSP/OLS and NPP/VIIRS data by applying a mask gray matrix algorithm to match the DMSP and VIIRS datasets into a specific mask and used a power function to calibrate the data [30]. The fundamental idea behind these methods is identical in that they first attempt to find an overlapping year between DMSP/OLS and VIIRS and then apply a regression-based model to fit the data value into the 0–63 range, which is the original data value range for the NTL DMSP/OLS data. However, these mathematical models still face the data value issue that the relevant data cannot exceed the numerical range of the original data to create an appropriate time-series NTL dataset with a good contrast and accurately represent the relative NTL digital values.

On the other hand, in the past three decades, NTL data retrieved by satellites have been implemented for applications such as evaluating urbanization processes, estimating GDP, and monitoring disasters and conflicts [25,31], where NTL data have demonstrated a close relationship to the particular economic situation [19,20,21]. For example, statistical methods were widely adopted to discover the correlation between different measurements of nighttime light (either DMSP/OLS or VIIRS) and GDP, which reflect socioeconomic development trends. Therefore, nighttime light has been found to be positively correlated with GDP and GRP at different spatial scales over a long temporal span. Nevertheless, when using NTL data to evaluate socioeconomic dynamics, many studies combined the NTL data to analyze a single measurement indicator between lights and a socioeconomic parameter [27,32]. Ge et al. used a linear regression model of bright area index and GDP to estimate the GDP of prefecture-level cities in Jiangsu Province [27]. Han et al. (2022) studied and analyzed the relationship between NTL and GDP and other indicators in China and India and found that NTL can reflect GDP and other factors under different conditions [28].

Common to all these NTL data-based studies, is a sole scale either on a national scale [33] or a municipal scale [34,35], even though they included two types of NTL data to extend the temporal length of study. Minimal research has been carried out on multi-scales to illustrate the changing relationship between NTL and economy from bottom to top, which could more precisely explain the long-term regional development. In addition, most studies considered the correlation coefficient as the most significant indicator to explore the relationship between NTL data and socioeconomic factors [36]. However, correlation coefficients cannot reflect multiple relationships, which could be important in the analysis process for NTL, GDP, GRP, and many other possible indicators, hence they cannot fully provide useful information among NTL data and other socioeconomic factors.

To address all the above-mentioned issues, we first designed an advanced data calibration framework to implement intra-calibration for DMSP/OLS data and inter-calibration for DMSP/OLS and VIIRS data. The proposed bilateral filtering method and tree prediction model were applied to resolving the conventional continuity and supersaturation problems, which have already demonstrated their powerful usage in solving data fitting problems [37]. Then we divided the calibrated data into three levels: metropolitan-level, central city-level, and suburban-level, and then illustrated and analyzed the NTL data and socioeconomic data on a multi-scale to explore the diverse development processes in different areas at multiple scales. Finally, multi-scaled analysis was implemented by applying NTL data and other socioeconomic factors to feed into a multivariable regression model, which could accurately predict a city and suburban area’s economy and social development in the future. The reminder of this paper is structured as follows: describing the study area and research data (Section 2), providing methodological details on the data calibration and data analysis process (Section 3), analyzing, and discussing the results (Section 4 and Section 5), and drawing conclusions (Section 6).

2. Study Area and Research Data

2.1. Study Area

The study area focused on Chongqing, a municipality in Southeastern China (Figure 1). The central urban area of Chongqing has an altitude lying mostly between 168 and 400 m, with mountains covering over 75% of its land area. With a total land area of 82,400 km², Chongqing has jurisdiction over 26 districts, eight counties, and four autonomous counties. Based on the 2020 census data, the population of Chongqing has reached 3,205,200. Chongqing is the first national strategic hinterland city and a leading economic development powerhouse upstream of the Yangtze River. Because of its unique geographical location and important role in the country, Chongqing had been experiencing rapid economic growth from 1992 to 2020.

As an essential modern manufacturing base, metropolitan Chongqing’s primary urban area is located in the west-central part of the city and comprises nine districts: Yuzhong, Dadukou, Jiangbei, Shapingba, Jiulongpo, Nanan, Beibei, Yubei, and Banan, all being the primary development sub-regions. The second tier for the city development is the one-hour economic belt around the metropolitan urban area, including Fuling, Changshou, Jiangjin, and other nine sub-urban districts, which is a potential development region under the policy of large-scale development of Chongqing. Northeast Chongqing and Southeast Chongqing are the two remaining areas with relatively slow economic development compared to the major urban area and one-hour belt. Northeast Chongqing consists of 11 suburban districts in the ‘Three Gorges Ecological Reservoir’ area and covers an area of 33,900 km². Therefore, it shoulders the important task of developing an “ecological economy.” Southeast Chongqing includes six suburban districts, which covers an area of about 19,800 km², with a forest coverage rate of nearly 50%. It is a minority inhabited area mainly by Tujia and Miao ethnic groups. The four hierarchical zones can reflect a unique big urban versus big rural area character; hence, they are ideal for carrying out a comprehensive multi-scale analysis with NTL data.

2.2. Used Data

Two types of NTL data were used in this study: DMSP/OLS Stable Nighttime Lights Annual Time Series (Version 4) and an SNPP/VIIRS Image (tile 3, 75N/60E), which can be downloaded from the official National Geophysical Data Center (NGDC) website of NOAA (https://www.ngdc.noaa.gov/eog/viirs/download_dnb_composites.html (accessed on 10 November 2022) for VIIRS and https://ngdc.noaa.gov/eog/download.html (accessed on 10 November 2022) for DMSP). The DMSP/OLS NTL data contained global images from 1992 to 2013 with a digital value (DN) range of 0–63, which is to fit the near infrared scanning band (0.4~1 μm) with the spectral resolution at six bits. The data were sourced from six sensors: F10 (1992–1993), F12 (1994–1996), F14 (1997–2003), F15 (2000–2007), F16 (2004–2009), and F18 (2010–2013), where F+number indicate different sensor IDs. The DMSP NTL data products used in this study are cloud-free images produced from averaging the original scanning value to derive the average visible, stable lights, and cloud-free coverages (https://ngdc.noaa.gov/eog/dmsp/downloadV4composites.html (accessed on 7 January 2023)). We extracted the time-series NTL data of Chongqing from 1992 to 2013 as part of the long-term sequential time -series NTL data using ArcGIS^®, and Figure 2a shows the total sum DN (TDN) value of the entire study area. The annual average masked image of NPP/VIIRS is used as another research data over the period of 2012 to 2020. The DN values of VIIRS were radiometrically calibrated to range from 0 to 254 (32-bit signed floating-point resolution for the Visible Infrared Imaging Radiometer Suite). The VIIRS sensor has a DNB (Day Night Band), which conducts the nighttime light observations of the earth surface at a resolution of 500 m (a six-fold increase over the original OLS). The VIIR data are georeferenced the annual composite night light product derived from daily scanning data in TIFF format exclusive of the effects of parasitic light, flashing, moon reflection, and cloud (https://eogdata.mines.edu/wwwdata/viirs_products/dnb_composites/v10/README_dnb_composites_v1.tx (accessed on 7 January 2023)). Figure 2b shows the TDN value of subset VIIRS images of Chongqing. The two figures demonstrate a significant difference between the original VIIRS data and the DMSP/OLS data in use. Therefore, it is necessary to create a unique long-term sequential time-series NTL dataset containing both VIIRS and DMSP/OLS data.

Due to data inaccessibility and incompleteness, only six socioeconomic factors that may affect the socioeconomic situation were selected and combined with the NTL datasets to describe the city’s development. The six factors are (1) the city’s gross domestic products (GDP, unit: 100 million CNY (Chinese Yuan)) at city scale and 10 thousand CNY at other scales, (2) the gross industrial output value (GIV, unit: 100 million CNY at city scale and 10 thousand CNY at other scales), (3) the added value of the secondary sector (AVSS, unit: 100 million CNY at city scale and 10 thousand CNY at other scales), (4) the added value of the tertiary sector (AVTS, unit: 100 million CNY at city scale and 10 thousand CNY at other scales), (5) total population (TP, unit: 10,000 people at all scales), and (6) the floor space of buildings completed (FSBC, unit: 10,000 m² at all scales). The time frame for these data is from 1999 to 2019 (The details of these six factors can be found in Supplementary Table S1 at three levels). The co-linearity may exist among these socioeconomic factors, however, due to the data accessibility and the method we applied, this co-linearity won’t affect our research result and goal.

3. Methodology

3.1. NTL Data Calibration Process

The calibration process is illustrated in Figure 3 with two major parts: DMSP/OLS data calibration and VIIRS data calibration. The first part follows Li et al.’s framework [38] with an update to address the supersaturation problem. Three processes are performed: inter-calibration among sensors, continuity calibration, and supersaturation calibration. The first process minimizes the sensors’ differences, makes the DNs as continuous numerical numbers, and most importantly, extends the DMSP/OLS data value to a more reasonable range. The VIIRS data calibration process includes an inter-calibration between DMSP/OLS and VIIRS, outlier calibration, and a four sub-step calibration for the non-overlapping years. The purpose of the first step is to provide a reference dataset for fitting the VIIR data value into the calibrated DMSP/OLS data value range by using a tree-based machine learning algorithm. After the error data have been corrected with the reference calibrated data, all the remaining VIIRS data are transformed into that specific value range to make the entire NTL dataset more generative. After this calibration process, the DMSP/OLS data from 1999 to 2013 and the VIIRS data from 2014 to 2020 are combined to create the unique long-term sequential time-series NTL data.

(1)

DMSP/OLS data calibration

As the raw light data were not calibrated for satellite radiation and the satellite sensor had been replaced, they must be calibrated to eliminate problems such as supersaturation of the central city, differences between temporally overlapping data, and the disappearance of luminous pixels in the preceding year [39]. Therefore, the DMSP/OLS data were calibrated in a number of steps. The DMSP/OLS images were first subset to delimit the study area using the Chongqing’s 2020 official division vector map. Then, the nearest neighbor algorithm was adopted to resample the subset images to a resolution of 1 km × 1 km to match the cell size of VIIRS data. This resolution was not significantly different from the original image resolution, hence avoiding problems like loss of numerical accuracy. During resampling, CGCS KG Zone 18 was adopted as the projection coordinate system. An intersection analysis with datasets of other years was executed using the 2013 image as the master image [40,41,42]. More details for the DMSP/OLS data calibration process could be found in Li et al.’s 2019 work [38].

(2)

VIIRS data calibration

The purpose of calibrating VIIRS data was to unify them into the same numerical range as that of the DMSP/OLS data so as to merge the two NTL datasets into one continuous data series spanning from 1992 to 2020. The calibration followed the same subsetting and resampling steps as with the DMSP/OLS data, and the calibrated data were sampled to 1 km × 1 km, during which the data value was unified as an 8-bit unsigned integer. A decision tree model was used to transform the VIIRS images of overlapping years into DMSP/OLS images, in which the overlapping DMSP/OLS images were used as the source data to transform the VIIRS images of non-overlapping years. In addition, outlier calibration and supersaturation calibration were also carried out.

Overlapping datainter-calibration: The 2012 and 2013 VIIRS and DMSP/OLS data overlapped and were used as the benchmark. In the past they had been fit together using a power function model without validation [43]. Instead, this paper proposed a regression-tree method based on decision tree as the dependent data are continuous. Compared with the traditional mathematical method, decision tree has some advantages, such as keeping the numerical range unchanged, a faster training speed, and more reliable accuracy [44]. It presents a mapping relationship between the value of an object and its attributes. Every tree node is an object, every tree route indicates a possible attribute value, and every route from the root node to the life node reflects a decision test. Figure 4 briefly illustrates how the model works. The original VIIRS NTL data of 2012 and 2013 were used as the input data and the overlapping DMSP/OLS NTL data were treated as the expected results. For the entire study area, about 22,000 cells had the same DN value between 1999 and 2013, and 70% of these stable NTL cell pairs were used to train the decision tree regression model; the rest were used as the testing data. In addition, to balance model accuracy and the generalization capability of the data calibration model, the hyper-parameters were optimized by setting the min_samples_leaf to 1; the min_samples_split to 2 after serval tests with the highest model accuracy.

Outliers calibration: Due to the significant differences between VIIRS and DMSP/OLS images, the data after fitting had many outliers, so it was necessary to address them using threshold calibration. After regression, the abnormal zero and non-zero values should be consistent with those of the target year. The regression-fitted target year was treated as the reference data, the images after the regression process were calibrated using Equation (1), where m stands for the reference year, $D{N}_{\left(n,i,j\right)}$ stands for the outlier digital value at location $\left(i, j\right)$, $D{N}_{\left(m,i,j\right)}$ is the digital value at the same location but in the reference year m.

$$D{N}_{\left(n,i,j\right)}=\left\{\begin{array}{ll}0& \mathrm{if} D{N}_{\left(m,i,j\right)}=0\\ D{N}_{\left(m,i,j\right)}& \mathrm{if} D{N}_{\left(n,i,j\right)}=0\end{array}\right.$$

(1)

Non-overlapping year calibration: After overlapping fitting and outlier modification, the calibrated VIIRS data with higher testing accuracy and the fitted decision tree regression model were used to adjust the remaining VIIRS data of non-overlapping years. This was carried out using the Gaussian filtering method [23]. In this study, it was replaced by bilateral filtering as it had the most significant advantage of not blurring the image boundary too much, a feature practically significant for geographic images [45]. The new pixel values after bilateral filtering were calculated using Equation (2), where w is the weight function, i and j, k and l, indicate the coordinate pairs for pixels, DN_D is the pixel value, σ_d is the distance parameter, and σ_r is the pixel difference parameter. They were set to 2 and 60, respectively, according to the estimation. In short, the calibration of non-overlapping years followed all the steps mentioned above in a four-sub-step calibration process, including tree regression, outlier calibration, bilateral filtering, and supersaturation calibration. The outlier calibration was included because some outliers were generated when predicting data for the overlapping years, and the purpose of carrying out supersaturation calibration was to eliminate the excessive concentration of high-value pixels and broaden the total DN value range.

$$\begin{array}{c}w\left(i,j,k,l\right)=\exp \left(-\frac{{(i-k)}^2+{(j-1)}^2}{2{\sigma }_d^2}-\frac{\parallel DN\left(i,j\right)-DN\left(k,l\right){\parallel }^2}{2{\sigma }_r^2}\right)\\ D{N}_D\left(i,j\right)=\frac{{{\displaystyle \sum }}_{k,l}DN\left(k,l\right)w\left(i,j,k,l\right)}{{{\displaystyle \sum }}_{k,l}w\left(i,j,k,l\right)}\end{array}$$

(2)

3.2. Spatiotemporal Analysis

To facilitate multi-scale spatiotemporal analysis, the entire study area was divided into three scales: entire Chongqing, metropolitan-level (including the metropolitan area, one-hour belt zone, Yudongbei zone, and Yudongnan Zone), and the suburban level of 38 districts. The six socioeconomic factors and NTL data from 1999 to 2020 were then correspondingly fed into matched-scale decision tree regression models to explore the relationship between socioeconomic factors and NTL data of sum, mean, and standard deviation, respectively. Therefore, the total modeling results are 3 × 1 (entire city) + 3 × 4 (4 sub-city regions) + 3 × 38 (38 suburban districts), or, 129 relationships to be established. Figure 5 illustrates the whole spatiotemporal analysis process:

With traditional regression analysis, the correlation coefficient could not fully explain how the different independent variables affected the dependent variable, so it can only interpret a one-to-one relationship [46]. Therefore, multiple regression analysis is needed to capture the multi-relationships. The decision tree was unsuitable for continuous value regression because it cannot predict a value larger than the original value, but it can calculate the influence of each independent variable (served by the six economic factors) on the dependent variable (served by the light index expressed as sum, mean, and standard deviation). In this part, we defined the importance of different features as the proportion of the average gain across all split-feeding features in the model.

4. Results

4.1. NTL Data Calibration Results

The regression parameters after inter-calibration among the DMSP/OLS sensed data are presented in

Table 1. Values of coefficients a, b, and c of data calibration.

Sensor	Year	a	b	c	R2
F10	1992	−0.0031	1.2854	0.5164	0.7972
	1993	−0.0014	1.1634	1.3032	0.8045
	1994	0.0022	0.9531	0.8577	0.8198
F12	1994	0.0021	0.9781	1.9929	0.7982
	1995	0.0065	0.6991	3.9826	0.8016
	1996	0.0093	0.5627	4.2085	0.8156
	1997	0.0104	0.4209	6.9318	0.8104
	1998	0.0118	0.4764	6.2895	0.8139
	1999	0.0082	0.6057	3.9276	0.8348
F14	1997	0.0054	0.7886	2.5985	0.8268
	1998	0.0061	0.7543	2.9812	0.7986
	1999	0.0009	1.0246	2.9390	0.8420
	2000	0.0054	0.7861	9.0337	0.8541
	2001	0.0003	1.2143	1.1397	0.8489
	2002	0.0000	1.2377	0.7930	0.8103
	2003	−0.0032	1.4156	0.9545	0.8802
F15	2000	0.0075	0.6452	3.7070	0.8125
	2001	0.0098	0.5019	1.4560	0.8520
	2002	0.0000	0.9901	2.4539	0.8379
	2003	−0.0039	1.3917	1.8164	0.8357
	2004	−0.0084	1.8061	0.9048	0.8298
	2005	−0.0051	1.5328	0.9462	0.8298
	2006	−0.0042	1.6072	1.0359	0.8039
	2007	−0.0089	1.8307	0.6289	0.7961
F16	2004	−0.0014	1.1668	1.0703	0.8679
	2005	−0.0028	1.4042	0.1650	0.9102
	2006	−0.0051	1.5113	0.0195	0.9348
	2007	−0.0000	0.0000	0.0000	0.0000
	2008	0.0056	0.6757	1.8309	0.9242
	2009	0.0071	0.5859	2.9597	0.9251
F18	2010	0.0085	0.4396	306207	0.8964
	2011	0.0069	0.4996	3.9897	0.7945
	2012	0.0085	0.3987	4.0598	0.8223
	2013	0.0074	0.4833	2.8755	0.8031

with all R² values greater than 0.79. This process weakened the variations among different DMSP/OLS sensors and the characteristic that high DN values are excessively concentrated in urban areas. After the calculation via the regression process, the maximum value of the DMSP/OLS data becomes 78 instead of 63. This value is fitted into VIIRS, whose original numerical range is 0 to 132. For the VIIRS data, which is calibrated using the decision tree, the model’s training accuracy was 0.9796 in 2012 and 0.9923 in 2013, with the testing accuracy being 0.9533 in 2012 and 0.9854 in 2013, demonstrating the tight harmonization between the calibrated DMSP/OLS and VIIRS based on the decision tree regression data fitting model. Figure 6 shows the 2020 VIIRS images before and after calibration. The range of the NTL DN value has been changed from 0–132 to 0–78, and both the contrast and saturation of the NTL DN become much more robust, demonstrating the successful conversion from VIIRS DN to continuous and uniform NTL data series.

The fractured and unconnected segments representing the raw NTL data now show up as a smooth and unbroken line (Figure 7). Starting with the minimal DN sum value of 47,617, the calibrated data increased steadily from 1992 to 2005, remaining relatively stable between 2005 and 2010. It had a slight decline in 2007 when the value dropped from 212,327 to 201,520, and then it bounced back to an upward trend from 2010 to 2020, which precisely matched the trend of the raw data. Between 2007 and 2010, there is a slight fluctuation due to the DMSP/OLS sensor problem [39], which is also reflected by the calibrated data. After that, the NTL TDN reached the maximum value of 347,233 in 2020 with the same upward swing as the original VIIRS data. In general, the NTL data calibration process enabled the two types of data sources to be unified and the integrated data retained the same trend as the individual DMSP and VIIRS datasets.

4.2. NTL Decision Tree Regression Results

Dissimilar to previous studies that only considered one NTL value to reveal the socioeconomic situation at a single scale, this study used three NTL indices of sum, mean, and standard and six socioeconomic indicators to explore their correlation at three scales (city—sub-city region—suburban district) by applying the multi-scale decision tree regression model. At the broadest scale, the highest DTR model accuracy is 0.95 between NTL mean and the indicators. At the sub-city region, the average DTR model accuracy is 0.80, 0.90, 0.85 for the NTL sum, mean, and standard, respectively. Therefore, mean NTL should be used. And at the suburban district-level, these average accuracy values become 0.84, 0.86, and 0.77. All these relatively high model accuracy scores demonstrate that DTR is able to establish accurate relationships between NTL mean and socioeconomic factors in this study.

After model validation, the F score is calculated to measure the influence degree of each independent variable on the dependent variable, which is a metric result that sums up how many times each feature is split with the DTR to describe feature importance. A higher F score indicates a stronger factor contribution to the regression fitting process, and only statistically significant results will be shown. The results at the city and sub-city regional scales are fully presented below, together with the results for some typical suburban districts, though all results can be found in the Supplementary File S1. Figure 8 illustrates the factor importance at the city scale. TP is the most significant factor correlated to the NTL indices, with an F score of 80, 222, and 331. Furthermore, it is the only factor statistically significant to the NTL sum. For both NTL mean and std, FCBC is the second most important factor. However, it has a much higher mean F score of 55 than that of the standard of only 3.

However, differences are remarkably obvious when the results among the sub-city regions are compared (Figure 9). The metropolitan-level results are very similar to those of the city scale in that TP is considered the only significant factor to be correlated with the NTL indices. The F score for sum increased to 96, but the F score for mean and the standard deviation dropped to 295 and 199, respectively. For the one-hour belt zone, the F score of TP is still high, but GDP is also treated as one critical factor in influencing the NTL indexes, especially for the standard deviation. The F score of GDP (116) shows that it is almost equally important to TP (122). The results of Northeast Chongqing still show a similar trend: TP is the dominant factor correlated to the NTL indices, and GDP is the second most important factor. However, for the NTL mean in Northeast Chongqing, GIV and FSBS also play a relative critical role, with an F score of 88 and 42, respectively. And for the standard deviation, GIV also slightly influences the NTL standard index. In Southeast Chongqing, TP still has the highest F score for all three NTL indexes. Nevertheless, the second most important factor changed from GDP to FSBC, which may indicate that the NTL indexes in this area are more closely related to real urban construction rather than the economic development process. After FSBC, the relatively essential factors are GIV and GDP, and AVSS first shows up with the DRT model that it would influence the NTL mean in Southeast Chongqing, with an F score of 54.

For the smallest scale, there are 38 suburban districts. Therefore, we run the model 114 (3 × 38) times and count the frequency and rank to generate the results in

Table 2. DTR factor results for the 38 suburban districts.

NTL Sum	1st Order	2nd Order	3rd Order	Other	Frequency	Average F Score
TP	30	7	1	0	38	145
GDP	8	16	6	2	32	68
GIV	0	1	8	12	21	24
AVSS	0	0	1	3	4	17
AVTS	0	0	1	3	4	10
FSCB	0	10	12	5	27	40
Total	38	34	29	25	126
NTL Mean	1st Order	2nd Order	3rd Order	Other	Frequency	Average Score
TP	34	3	1	0	38	282
GDP	4	14	9	6	33	104
GIV	0	3	5	17	25	64
AVSS	0	0	2	9	11	18
AVTS	0	0	3	5	8	22
FSCB	0	16	13	3	32	92
Total	38	36	33	40	147
NTL STD	1st Order	2nd Order	3rd Order	Other	Frequency	Average Score
TP	27	3	0	8	38	164
GDP	11	16	3	0	30	96
GIV	0	1	12	12	25	23
AVSS	0	2	2	5	9	37
AVTS	0	0	2	1	3	21
FSCB	0	7	11	7	25	43
Total	38	29	30	33	130

. It provides information on the dominant relationship between NTL indexes and socioeconomic factors. As the table reveals, TP, GDP, and FSCB are the most important factors based on their close correlation with all the three NTL indexes. However, there are still differences between them. For example, the total frequency number for NTL mean is higher than the other two, which means that more factors have been involved in the DRT process to affect the relationship. In addition, FSCB is the second closely correlated factor to NTL mean, with the highest number for the second and third ranks for all the 38 suburban districts, where it is usually the third or lower ranked factor. And for NTL standard, the dominance of TP was challenged by GDP in that 11 suburban districts have GDP as the factor most closely correlated to NTL standard. Additionally, 16 districts have it as the second most crucial factor in influencing NTL standard.

Four typical suburban districts were also selected for detailed analysis to reveal their development process. Figure 10 shows that two of them belong to the metropolitan area and one-hour belt zone. They have fewer significant factors than the other two areas. Additionally, TP and FSCB are usually the two most important factors to be correlated with the NTL indexes, indicating that construction and city population indeed affected the NTL value in these areas. Moreover, the number of factors to be significantly correlated to the NTL indexes increased for Northeast and Southeast Chongqing, such as TP, GDP, AVTS, GIV, and FSCB all impacting the NTL values. So, the development in those areas differs from the metropolitan area and the one-hour belt zone.

4.3. Multi-Scale Analysis of NTL Data

The tight correlations between NTL indexes and specific socioeconomic factors have been demonstrated through the DTR model in the previous section. For example, NTL sum is significantly correlated toTP, which could indicate the integrated intensity of urban development, because a higher TP means that the city is more attractive to people, and hence draws a larger workforce to the city. More laborers will create more intensive and integrated development to satisfy the growth needs and make the city more attractive, leading to a virtuous cycle [47,48]. As illustrated in Figure 11, the NTL sum of the entire Chongqing keeps increasing, from nearly 100,000 to almost 350,000. This increase reflects an increased population from 1999 to 2020, indicating a significant development of the entire city. At the sub-city region level, the metropolitan area and the one-hour belt have similar increasing trends until 2012. Afterwards, the one-hour belt zone has a slightly higher increase than the metropolitan area, which reflects a faster development of the suburbs around the metropolitan area, matching the development priority based on the city council’s plan. On the other hand, NTL sum of Southeast Chongqing has stayed relatively stable, suggesting less intensive integrated development or even bare development. This also matches the developing status for that specific area as the government investment in that region is the lowest of the four sub-city regions because its mountainous terrain severely restricts development.

NTL mean is calculated using the NTL sum divided by the total area of a specific zone and has the highest number (147) of correlated factors on the suburban-district scale. Therefore, this value could represent the average development intensity for each cell inside the zone [49]. As Figure 12 illustrates, all the mean values on both the city scale and sub-city region scale have a similar increasing trend from 1999 to 2020. Both the one-hour belt zone and metropolitan area have a higher mean value than the entire city (from 1.24 to 4.20) at every annual increment, which reveals that the development intensity in these two zones is higher than the average level of the entire Chongqing. For the metropolitan area, in particular, its mean value increased from 3.0 to 8.35, almost twice as high as that of the entire city, indicating a powerful development intensity and reflecting the development priority. In contrast, northeast and southeast Chongqing have much lower values than the entire city, and all the NTL mean values are below 2.0, meaning that the latter two zones’ development intensity is relatively low, and there is a very limited urbanization process inside these two zones.

As shown in Table 1, the NTL standard value is most influenced by TP, GDP, GIV, and FSCB, which reveals a more general socioeconomic situation. The value of standard represents the difference between one NTL cell value to the mean NTL value inside a given zone. A higher standard value means a more prominent contrast among cells, and a lower value means a more uniform value inside this zone. Therefore, it can reflect the disparity between urban and rural areas [24]. For example, a zone having a very low standard is either well urbanized or completely rural without development. From Figure 13, we can see that the NTL standard for entire Chongqing keeps increasing steadily from 1999 to 2020, with the most significant increase during 2002–2006, from 5.98 to 7.89, which indicates that disparity between urban and rural areas in Chongqing steadily widened. The one-hour belt zone has a similar trend as entire Chongqing, but it has a lower standard value. For Northeast Chongqing and Southeast Chongqing, the NTL standard also keeps rising slightly, but the absolute value is much lower than that of the entire city, which reflects the relatively steady development status for these two sub-city regions. Furthermore, the rising trend for the metropolitan area is the most obvious, with much higher standard values over the study period. The standard value in 1999 was 13.62 and jumped to 23.44 in 2020. This abrupt increase reveals that the disparity between urban and rural areas inside the metropolitan area is increasingly enlarged and that many rural areas have been transformed into urban land.

Figure 14 shows that the NTL sum curves all indicate an increasing trend. In 1999, Wanzhou in Northeast Chongqing had a higher NTL sum than Bishan within the one-hour belt. And its value almost catches up the value of Jiulongpo in the metropolitan area. However, it was surpassed by Bishan in 2013 because it had only a relatively low increase. Bishan’s NTL sum value was more than 3000 lower than Jiulongpo’s in 1999, however, the gap was narrowed to less than 300 by 2020. This could be explained by the fact that the metropolitan area had stable development, but the development in the surrounding suburbs was rapid. Therefore, the increase of the NTL in the later period mainly comes from newly developed areas such as Bishan County. Wulong always had a relatively low NTL sum, indicating a weakly integrated development intensity.

Unlike NTL sum that describe the integrated development intensity for the urban area, the NTL mean can reflect the average development intensity for a specific zone. Figure 15 provides the information on NTL mean for the four typical suburban district zones. Of these zones, Jiulongpo has a mean value remarkably higher than others, suggesting a stronger average development for the entire zone. The rest of the four zones were very close to each other at the beginning, and the gap between Wanzhou and Wulong is quite stable, meaning a relatively stable development speed for these two zones. However, the mean value of Bishan jumped to over 15 and quadrupled in 20 years, which indicates a very fast development speed of the one-hour belt region.

Figure 16 illustrates the NTL standard value for these four typical zones. It shows that Bishan has an increasing trend from 4.5 to 13.5, which indicates a fast development of this district, and the disparity between urban and rural areas is also enlarged. The value of Jiulongpo fluctuated during the period. It dropped from 16.3 to less than 16 and then remained stable for almost ten years. Then it reached the highest value of 18.3 and decreased slightly to over 16. The overall flat curve of Jiulongpo reveals that the disparity is very limited inside the metropolitan area. This could also be explained by the fact that the rural area is relatively small and the development in that zone mainly takes the form of urban upgrading/renewal instead of new constructions. The standard values of Wanzhou and Wulong both have a stable increase trend, different from that of Bishan, which also shows the increased disparity between urban and rural areas, but not as obviously as that of Bishan because of the lower development speed.

5. Discussion

This study generated a continuous time sequential NTL dataset by fusing DMSP/OLS and VIIR NTL data and applied the calibrated new NTL dataset to reveal the relationships between NTL and socioeconomic factors at multi-scales. Comparing to previous studies, the supersaturation issue for the NTL data has been addressed and the range for digital value has been adjusted to 0–80 (instead on only 0–63 or 0–132) through the ML-based process. This value range enhanced the contrast of the NTL (Figure 6), hence making a more smooth and robust time sequential NTL dataset of 30 years. The data calibration accuracy between the two types of data is rather high at 0.95–0.98 (R²), which is higher than that of conventional calibration methods [23,25]. Moreover, the in-calibrated data could also be put into the model to demonstrate the proposed calibration method in our future studies.

Next, through the DTR model, we tried to find out the relationship between NTL indexes (sum, mean, standard) and socioeconomic variables. For the entire study area, the highest R² score is 0.95 between NTL mean and other variables, which is much higher than many previous studies with conventional ordinary regression that produced an R² around 0.85 to 0.90 [9,10,11,19,20,30]. Although the R² value drops to 0.86 at the micro-scale, it is still slightly higher than the 0.8 obtained in other studies using only one or two socioeconomic variables at the same scale [27,32,47]. In addition, the accuracy of DTR prediction also varies with scale in that the closeness of the relationship changes from the macro-scale to the micro-scale. Specifically, the relationship of NTL sum and independent variables becomes closer from less than 0.7 to 0.84 and that of NTL mean and standard becomes looser from 0.95 to 0.86, and 0.89 to 0.76, respectively (Figure 17). From these results, we can conclude that NTL mean should be the first choice for predicting socioeconomic situation at the city and sub-city scales, followed by NTL standard. At the suburban district scale, the combination of NTL mean and sum could be a good indicator. The prediction result for NTL standard is relatively low at the suburban districts. However, for certain areas it has a quite high value of over 0.95 (Fengdu: 0.968, Xiushan: 0.961, Youyang: 0.973). These three areas have one thing in common: they have relatively concentrated middle-range urban areas and scenic view areas, which reveals that NTL standard may also be used to indicate some specific areas at the suburban scale. These results provide more information on matching NTL indices with appropriate scale studies than most related studies that used only one NTL index [32,38,45].

Furthermore, the number of significant factors relating to the NTL indexes also changes as the modeling scale varies. More socioeconomic factors contribute to the correlation as the scale becomes more local. For the entire city, only one or two variables are important to the relationship, but the number increases to three to five at the suburban district level. For example, at the largest scale, only TP is significant to the NTL indexes but TP, GDP, AVIS, GIV, and FSCB all make contributes to the NTL indexes at the micro-scale. In addition, the factor importance values also change as the scale changes. These results all indicate that the NTL indexes could explain more complicated situations as they scale down. Therefore, the multi-scaled process is able to reveal the relationship between socioeconomic factors and NTL indexes more reasonably and indicate the factor importance variation at different scopes, which could not be fulfilled by single-scale study.

In addition, this study can be improved by taking aerosol into consideration. Atmospheric particulate matter greatly influences satellite observations and the OLS/DMSP and VIIR light results are no exceptions [50,51]. Even though the study area is not a heavily polluted city, it must be acknowledged that neglecting the effects of aerosols could adversely affect the calibration process, which should be explicitly considered in a future study.

6. Conclusions

In general, this study successfully constructed a continuous time NTL dataset and reveal the relationships between NTL indices and socioeconomic factors using a DTR model that is run more than one hundred times at three scales of entire city, sub-city region, and suburban district. All the model results from different scales demonstrate a reasonable correlation between NTL indexes and socioeconomic factors, with an average model accuracy over 0.85. The significance of our research lies in the introduction of machine learning methods to correct luminous images and the discovery of the relationship between the luminous image data, composed of DMSP/OLS and VIIRS, and socioeconomic factors in the DTR analysis. The presented results could be used to analyze and explain the comprehensive urban development process spatiotemporally. In addition, multi-scale analysis could provide more insightful information to government planners from different perspectives and enable them to understand the development with extensive knowledge. They will know how the city develops at different levels through the NTL changes, as which NTL indexes could be used to illustrate the development situation and which factors are the most significant in these development processes at a particular scale, from the whole city to a specific suburban district. And with such information, they could make more informed development plans in future.

However, even if the data type and the numerical range remain unchanged, outliers in the tree model still need to be corrected, so the calibrated results need to be corrected too. Researchers can consider adopting a decision tree for each part of the calibration of luminous images due to the advantages this study introduced in the future. In addition, this study focused on only one city, and hence the obtained results have a limited practical significance on the national scale. Therefore, the research area can be expanded to include more cities to obtain a robust calibration model and predictors to make the results more universally applicable in the future. Moreover, since atmospheric particulate matter greatly influences satellite observations, the analysis of the aerosols effect should be included in future studies.