Assessing Uneven Regional Development Using Nighttime Light Satellite Data and Machine Learning Methods: Evidence from County-Level Improved HDI in China

Abstract

Uneven regional development has long been a focal issue for both academia and policymakers, with numerous studies over the past decades actively engaging in discussions on measuring regional development disparities. Generally, most existing studies measure the Human Development Index (HDI) using relatively simple indicators, with a focus on national and provincial scales. As a crucial component of regional development, counties can directly reflect the regional characteristics of socio-economic progress. This study employs a multi-dimensional approach to develop an improved Human Development Index (improved HDI) system, using machine learning techniques to establish the relationship between nighttime light (NTL) data and the improved HDI. Subsequently, NTL data are utilized to infer the spatial distribution characteristics of the improved HDI across China’s county-level regions. The improved HDI for county-level areas in the Ningxia Hui Autonomous Region was validated using a machine learning model, resulting in a Pearson correlation coefficient of 0.93. The adjusted R-squared value for the linear fit was 0.86, and the residuals were relatively balanced, ensuring the accuracy of the simulations. This study reveals that 1439 county-level units, representing 50% of all county-level units in China, have development levels at or above the medium level. At the provincial and national levels, the improved HDI shows significant clustering, characterized by a multi-center pattern with declining diffusion. The spatial distribution of the improved Human Development Index remains closely associated with the natural geographic background and socio-economic development levels of the county regions. Lower HDI values are predominantly found in the inland areas of central and western China, often in ecologically sensitive areas, inter-provincial border zones, and mountainous regions of mainland China, sometimes forming contiguous distribution patterns. This underscores the need for the government and society to focus more on these specific geographic development areas, promoting continuous improvements in health, education, and living standards to achieve coordinated regional development.

1. Introduction

The exploration of regional uneven development originated in the Marxist political economy and has since been widely applied across various social science disciplines. The study of regional disparities is inherently interdisciplinary, encompassing fields such as economics, geography, sociology, and political science. Notably, there is a broad consensus among scholars and policymakers that regional uneven development is an intrinsic characteristic of the capitalist system [1].

Neoliberal perspectives posit that market liberalization and the free flow of capital and labor will lead to balanced regional development; however, this theory lacks empirical support and relies on unrealistic assumptions. In the mid-20th century, Myrdal [2] and Hirschman [3], nearly contemporaneously, introduced the theories of cumulative causation and polarization-trickle effects, respectively. Both scholars focused on how market forces contribute to regional development disparities, profoundly influencing subsequent regional development theories. During the 1960s and 1970s, the deepening of globalization and the international division of labor further exacerbated regional imbalances. Marxist scholars, such as David Harvey [4], Neil Smith [5], and Doreen Massey [6], systematically explored issues of uneven regional development. They integrated Marxist political economy with geography and urban studies, focusing on how competition between capital and labor drives the capitalist economy and influences regional disparities, thereby introducing a series of new theoretical perspectives and analytical frameworks.

By the late 20th century, Krugman’s New Economic Geography had a significant impact on regional economics, building on the work of scholars like Myrdal and Hirschman by considering the effects of economies of scale and transport costs on the uneven development of regions. Krugman’s theory provided new insights into regional development patterns, aligning with trends in globalization and economic integration [7]. In the early 21st century, theories emphasizing cultural and institutional analyses highlighted the influence of non-economic factors on regional development, such as the roles of social relationships, knowledge, innovation, and institutional capacity—factors that were not sufficiently addressed in Krugman’s New Economic Geography. Notably, the role of the state in promoting regional development cannot be overlooked, as governments worldwide strive to achieve balanced development across regions through various policy and planning interventions [8,9]. In conclusion, regional disparities have always been a complex and systemic issue, necessitating the integration of diverse theoretical perspectives, including New Economic Geography, cultural and institutional analyses, and Marxist political economy, to develop a more comprehensive understanding of uneven development across different regions.

In empirical research, scholars utilize diverse metrics to assess regional development imbalances, yet each metric exhibits certain limitations. Metrics, such as the GDP [10,11], HDI [12], Gini Coefficient [13], and Sustainable Development Indicators [14], have been employed in studies across various global regions. Meanwhile, empirical studies on these indicators typically involve large-scale data, with many analyses focused at the national or provincial level. While some studies integrate multiple indicators to mitigate the limitations of individual measures, challenges remain in elucidating the drivers of regional disparities. For instance, GDP captures the economic dimension of regional development but fails to directly reflect residents’ living standards or disposable income. The Gini Coefficient can assess income and wealth distribution inequalities, but it requires comprehensive income sequence data. Sustainable Development Indicators demand high-quality, complete datasets, leading to varying degrees of constraints in applying these metrics across different scales and regions. In 1990, the United Nations Development Programme (UNDP) created the Human Development Index (HDI). The HDI makes up for the limitations of GDP in measuring the level of national economic and social development of a country or region, which is mainly composed of three indicators: life expectancy, adult literacy rate, and per capita GDP [15]. However, the HDI also has its own limitations, such as a lack of diversity of factors affecting human development in the selection of the HDI indicators. In addition to the choice of the three most basic dimensions of lifespan, education, and income, factors such as security, human rights, politics, the depletion of natural resources, and the consideration of environmental conditions reflect the level of economic and social development of a region to a certain extent and are highly valued [16].

On this basis, the Inequality-adjusted Human Development Index (IHDI) came into being. The IHDI is a refinement of the HDI that accounts for inequality in the distribution of each dimension across the population. The IHDI adjusts the average value of each dimension according to its level of inequality by applying a “discount” for inequality. When there is no inequality among people, the IHDI value equals the HDI value; however, as inequality increases, the IHDI value falls below the HDI value. The IHDI is computed as a geometric mean of the inequality-adjusted dimensional indices and is based on a distribution-sensitive class of composite indices proposed by Foster, Lopez-Calva, and Szekely [17], drawing on the Atkinson [18] family of inequality measures. However, research on the Inequality-adjusted Human Development Index (IHDI) is constrained by several limitations, including its dependence on high-quality and comprehensive data, inadequate representation of inequalities related to gender and race, and significant correlations among various indicators. These challenges underscore the need for more objective methodologies to effectively capture the spatial heterogeneity of human development levels across different regions [12,19,20].

Utilizing more objective indicators to assess regional development levels has emerged as an urgent issue. The advancement of remote sensing technology offers new perspectives for addressing this challenge, particularly with the rapid development of nighttime light (NTL) remote sensing data [13]. NTL data have the advantage of not being affected by geographical boundaries and human factors [21], which can effectively avoid the problem of system errors caused by government misreports or different measurement methods, and can objectively reflect human social and economic activities [22,23]. Since the DMSP satellite of United States obtained nighttime light data in the 20th century, nighttime light data have been widely used in various scales in the fields of population spatialization, carbon emissions, agricultural economy, urban expansion monitoring, economic development and human habitation law research, which confirms the objective accuracy of nighttime light data used to characterize the level of regional economic development [24]. Existing studies on GDP analysis [22], population density [25], disaster prediction [26], and other areas demonstrate that variations in NTL intensity effectively simulate the temporal and spatial changes in human activities [27,28], which confirms the correlation between NTL and economic development. Several studies from developing countries have attempted to refine the Human Development Index (HDI) and have validated the strong spatial correlation between nighttime light data and the HDI [29]. Therefore, the combination of the nighttime light index and economic statistics can better measure the social development of a country or region [24,30].

A review of the existing literature highlights several research gaps: (1) The Human Development Index (HDI), despite being intuitive and surpassing single economic indicators, considers only three dimensions, life expectancy, education level, and per capita income, which limits its comprehensiveness in assessing human development. (2) The Inequality-adjusted Human Development Index (IHDI), introduced by the United Nations Development Programme (UNDP), has not been widely adopted or consistently applied. Its calculation across diverse countries and regions requires stringent consistency in statistical standards, accuracy, and timeliness. Integrating nighttime light remote sensing data with machine learning techniques offers a promising approach to address these limitations of the IHDI. Additionally, enhancing the HDI to capture a more holistic representation of regional and human development remains essential.

Based on the identified research gaps, this study integrates an improved Human Development Index (HDI) with nighttime light data, using counties in China as the study areas. The analysis investigates the significant disparities in the improved HDI at the county level across China, offering a novel perspective on achieving common prosperity and enhancing regional development coordination. The spatial heterogeneity of these disparities is examined through the lens of various special geographical regions, underscoring the advantages of nighttime light data in estimating the improved HDI at the county scale. Additionally, the study emphasizes the spatial variations in the improved HDI among Chinese counties. The second section of the paper outlines the study area, methodology, and data sources. The third section details the calculation of the improved HDI at the county scale using nighttime light data and machine learning techniques. The final section compares the study’s findings with existing research and explores the potential reasons for observed differences.

2. Materials and Methods

2.1. Study Area

Our study area encompasses the county-level administrative divisions across China, with the exception of those in Taiwan Province of China, due to the insufficiency of county-level data available for inclusion in our calculations. We chose counties in Shanxi, a province in central China, as the calculation area for the improved HDI to ensure that large biases were avoided when making county forecasts in China. We also selected counties in the Ningxia Autonomous Region as validation areas for nighttime light predictions. In addition, we conducted a comparative analysis of China’s national-level poverty-stricken counties and other non-poverty-stricken counties, classified by the China Poverty Alleviation Office. We also conducted a comparative analysis of inter-provincial borders and other non-interprovincial border counties in China. In this study, mountainous counties in China were also selected for comparative analysis with other non-mountainous counties. Through the analysis from the abovementioned different perspectives, it is illustrated that the improved HDI based on NTL inversion can show the imbalance in development among regions.

2.2. Method

The research on using NTL to invert socio-economic development is developing rapidly, but there are limited studies on the inversion of an improved HDI using NTL. At the same time, the rapid development of machine learning and data mining technology has greatly promoted the accuracy of the NTL inversion of other socio-economic factors. For example, Xu et al. (2021) used machine learning and NTL to simulate the poverty index of poverty-stricken counties, and Wan et al. (2023) used NTL to predict the HDI of China’s border counties, and the results were very accurate [29,31]. Inspired by these studies, we used machine learning methods to explore the inversion of NTL to the improved HDI. Machine learning methods have certain advantages over traditional methods and are often the most efficient and robust methods in supervised learning mechanisms [32].

We selected the corrected NPP nighttime light satellite data in 2020 as the prediction base data. In our model training, we selected nocturnal light features including the mean, standard deviation, and 90th percentile of nocturnal light intensity within each county. Additionally, we incorporated population density as a corrective variable in the training and prediction of the improved Human Development Index. This incorporation enhances the model’s predictive stability and accuracy. By selecting different characteristic variables of NTL, we calculated the improved HDI of Shanxi counties as the dependent variable to form different combinations and construct a machine learning regression model. Through the training of different models, the model with the highest R² was selected as the most predictive basis (model 2.16 in

Table 1. Fitting indices of different machine learning method models.

Model Number	Model Type	RMSE (Validation)	MSE (Validation)	R2 (Validation)	MAE (Validation)
1	Tree	0.064	0.004	0.634	0.046
2.1	Tree	0.064	0.004	0.634	0.046
2.2	Tree	0.074	0.006	0.502	0.054
2.3	Tree	0.085	0.007	0.348	0.061
2.7	SVM	0.069	0.005	0.573	0.046
2.8	SVM	0.067	0.004	0.598	0.049
2.9	SVM	0.070	0.005	0.565	0.055
2.12	Integration tree	0.063	0.004	0.648	0.044
2.13	Integration tree	0.061	0.004	0.662	0.043
2.14	Gaussian process regression	0.065	0.004	0.625	0.046
2.15	Gaussian process regression	0.063	0.004	0.646	0.044
2.16	Gaussian process regression	0.058	0.003	0.702	0.041
2.17	Gaussian process regression	0.061	0.004	0.666	0.043
2.18	Neural networks	7.134	50.889	0.520	0.745

). The accuracy of the regression model was also measured using the mean absolute error (MAE) and root mean squared error (RMSE).

$$RMSE=\sqrt{{\displaystyle \frac{\sum _{i=1}^n{(Y_i-X_i)}^2}{n}}}$$

$$MAE={\displaystyle \frac{\sum _{i=1}^n|Y_i-X_i|}{n}}$$

We performed regression predictions in MATLAB (R2024a). Based on the corresponding NTL characteristic variables extracted from the county vector boundary in China and model 2.16 obtained by the above training, the improved HDI of all counties in China was predicted and verified by combining it with the calculated improved HDI of the counties in the Ningxia Autonomous Region. The Pearson correlation coefficient of the two reached 0.93, the linear fitting R² adjustment value reached 0.86, and the residual distribution was relatively balanced. The verification method used was 5-fold cross-validation, and the model fitted well (Figure 1). Therefore, the overall stability of model 2.16 verified the accuracy of the simulation (Figure 2).

Gaussian Process Regression (GPR) is a probabilistic, non-parametric machine learning method used for regression tasks. It offers several advantages, such as providing confidence intervals for predictions and employing the maximization of the marginal log-likelihood function for model selection, which enhances objectivity and automation. Additionally, GPR allows the integration of prior knowledge into the kernel function, and the choice of kernel is highly flexible. These features make GPR a particularly suitable model for regression predictions.

Normality of Residuals: A Shapiro–Wilk test was performed to evaluate the normality of the residuals, which indicated that they were approximately normally distributed (p = 0.08). This result supports the assumptions underlying our regression model. Correlation between Residuals and Predicted Values: A scatter plot of residuals against the predicted improved HDI values (Figure 2b) shows no significant correlation (r = 0.05, p = 0.56), suggesting the absence of systematic bias within the model. Homoscedasticity of Residuals: Analysis of Figure 2c,d, along with the results of a Breusch–Pagan test (p = 0.21), confirmed that the residuals exhibit homoscedasticity, indicating consistent error variance across the range of predicted values. Visual Analysis of Residuals: Visual inspection of the residual plots shows no discernible patterns or trends, with residuals randomly distributed around zero. This further supports the appropriateness of the model and suggests that it adequately captures the variability in the data.

2.3. Data

2.3.1. Nighttime Light (NTL) Satellite Data

We selected the global nighttime light NPP-VIIRS-like data produced by Chen et al. (2020) as the base data for inverting the improved HDI. The data were based on the Suomi National Polar Orbit Partner Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) [33]. The overall accuracy of the data is guaranteed to a certain extent. Regarding the accuracy of the pixel scale, the R² of the dataset as a whole reached 0.87, and its RMSE value was 2.96. At the city scale, the R² of the dataset reached 0.95, and its RMSE value was 3024.62. We obtained this product from the following URL: https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/YGIVCD (accessed on 10 April 2024).

2.3.2. Administrative Boundary Data for Counties

Administrative boundary data were also the core basic data of this study. We used data from the 2021 administrative boundary data of China from the Resource and Environmental Science Data Platform of the Chinese Academy of Sciences (https://www.resdc.cn/DOI/DOI.aspx?DOIID=120, accessed on 10 April 2024). In our study, boundary data at the county scale in China were also used. Whether a county was an inter-provincial border county was determined based on the inter-provincial border of the dataset, and whether the county was a mountainous county was judged using the list of mountainous counties that Li et al. (2018) used [34] (https://geodoi.ac.cn/WebCn/doi.aspx?Id=1018, accessed on 10 April 2024). Whether a county was a poverty-stricken county was determined in accordance with the List of Key Counties for National Poverty Alleviation and Development Work, issued by the former National Leading Group for Poverty Alleviation and Development (http://politics.people.com.cn/n/2014/1017/c1026-25854065.html, accessed on 10 April 2024).

2.3.3. Other Relevant Data

We also obtained other relevant data from Shanxi Province and the Ningxia Hui Autonomous Region, such as the GDP per capita, urbanization rate, per capita added value of secondary and tertiary industries, years of schooling, etc. Such data were obtained through the statistical yearbooks of Shanxi Province and the Ningxia Hui Autonomous Region. Data such as illiteracy rates were obtained through statistical bulletins and data from the Seventh Population Census. All of these data were used to calculate the improved HDI.

3. Results

3.1. Weight Analysis

In order to ensure objectivity, we used SPSS 20 software to calculate the weights of the indicators. Among the ten selected indicators, the weights obtained were as follows (

Table 2. Specific parameters of the entropy method.

Variable	Item	Information Entropy (e)	Information Utility Value (d)	Weight (%)
X1	Urbanization rate	0.952	0.048	8.119
X2	GDP per capita	0.933	0.067	11.332
X3	Added value of the secondary industry per capita	0.907	0.093	15.676
X4	Added value of the tertiary industry per capita	0.85	0.15	25.299
X5	Number of enterprises above scale per 10,000 people	0.95	0.05	8.414
X6	Average years of education	0.959	0.041	6.862
X7	Number of primary and secondary school students per 10,000 people	0.987	0.013	2.259
X8	Disposable income per capita	0.962	0.038	6.419
X9	Number of hospital beds per 10,000 people	0.918	0.082	13.829
X10	Illiteracy rate	0.989	0.011	1.791

): urbanization rate, α₁ = 8.119% (X₁); per capita GDP, α₂ = 11.332% (X₂); per capita added value of secondary industry, α₃ = 15.676% (X₃); per capita added value of tertiary industry, α₄ = 25.299% (X₄); number of enterprises above the designated size per 10,000 people, α₅ = 8.414% (X₅); average years of education, α₆ = 6.862% (X₆); number of primary and secondary school students per 10,000 people, α₇ = 2.259% (X₇); per capita disposable income, α₈ = 6.419% (X₈); hospital beds per 10,000 people, α₉ = 13.829% (X₉); and illiteracy rate, α₁₀ = 1.791% (X₁₀).

3.2. Assessed Values of Improved HDI at Each County Level in Shanxi Province

The improved HDI values of the Shanxi counties were then divided into five development levels according to the Natural Breaks (Jenks) in ArcGIS 10.8. As can be seen in Figure 3, the range [0.0574–0.1188] indicates a low level of human development; the range [0.1189–0.1867] indicates a low-to-medium level of human development; the range [0.1868–0.2627] indicates an average level of human development; the range [0.2628–0.4089] indicates a medium-to-high level of human development; and the range [0.4090–0.6474] indicates a high level of human development.

In 2020, there were three counties in Shanxi with an improved HDI value of more than 0.409, which were the Yingze District, Xiaodian District, and Xinghualing District of Taiyuan City, in order of their development levels. The improved HDI results also show that 17 counties were at a medium-to-high level of human development, 27 counties were at a medium level of human development, 32 counties were at a low-to-medium level of human development, and 38 counties were at a low level of human development. Regarding the numbers of counties at different levels, counties with a lower middle level of human development were the majority, accounting for a large proportion, because Shanxi Province has a limited overall development level among all provinces in China. It is located in the central region, mainly relying on the coal industry as its pillar industry, and the level of social and economic development is limited. From the perspective of distribution, most of the high-improved HDI-value areas are located in the main municipal districts of key development cities in the province, which are rich in resources, have a good geographical location, and are supported by policy funds. These conditions laid a certain foundation for the agglomeration of elements and the regional division of labor and cooperation.

To a certain extent, the development levels of the counties appear to be polycentric, decreasing from the center to the periphery. Counties with higher development levels can drive the development of surrounding counties with lower development levels. As can be seen in the figure, most of the counties bordering those with a high level of human development have an above-average level of human development. Similarly, the counties around those with a medium-to-high level of human development have medium and low-to-medium human development levels. This is because under the radiation and driving effects of the core county’s economy, the surrounding counties strengthen economic cooperation within the core area, and their economic strength is also improved to a certain extent. The business objects of core county industries need to have certain basic economic conditions so as to realize the application and utilization of more spillover factors in core counties [35]. Thus, the two coordinate with and promote each other, improve production efficiency, and form a decreasing improved HDI development pattern.

3.3. Spatial Pattern of Improved HDI at the County Level in China

Based on model 2.16 shown in the Section 2.2 and the improved HDI of each county in Shanxi Province, we predicted the improved HDI of counties in China and performed spatial mapping in ArcGIS 10.8 software. Using the Natural Breaks (Jenks) in ArcGIS 10.8, we divided the results into five levels of development (Figure 4): The range [0.0668–0.0931] indicates low human development, [0.0932–0.1492] indicates low-to-medium human development, [0.1493–0.2221] indicates medium human development, [0.2222–0.2781] indicates medium-to-high human development, and [0.2782–0.4381] indicates high human development.

Among the counties, there are 201 with a high level of human development, and their distribution is relatively scattered. These counties are mainly distributed in China’s coastal areas, especially in the three national urban agglomerations, including the Beijing–Tianjin–Hebei urban agglomeration, the Yangtze River Delta urban agglomeration, and the Pearl River Delta urban agglomeration. These counties have better location conditions, superior natural environmental conditions, and a higher level of economic and social development, so the improved HDI values of these counties are higher. In addition, there are 1439 counties with a medium or higher development level, accounting for 50% of the total number of counties in China, mostly distributed in the eastern and central parts of China. There are 559 counties with low-to-medium human development and 880 counties with low human development, mostly located in the central, western, and northeastern parts of China.

In general, the improved HDI values of counties in China decrease from the east to the west, and the values in the south are greater than those in the north. Despite the counties within urban agglomerations having developed better socio-economically under the urban agglomeration policy, there are regional differences. The urban agglomeration index is higher in the eastern region because it has been the focus of China’s national development strategy since the reform and opening up of China. Cities such as the Chengdu–Chongqing urban agglomeration and Xi’an in the western region are due to the very important foundation laid by the third-line construction after the founding of the Peoples Republic of China, so the improved HDI of these regions are higher compared to other regions [36,37]. Although counties with low-to-medium human development still account for half of the total number of counties in China, most of them are in the transition stage to the medium development level, and the situation is improving. Due to the influence of geographical locations, traffic conditions, industrial characteristics, and other conditions, the areas with low improved HDI values are distributed in the inland areas of the central and western regions and the underdeveloped areas of the northeast. However, the areas with high improved HDI levels are mostly distributed in the main municipal districts of provincial capitals and their surrounding cities and are mostly in the form of clusters. Isolated high-value areas are rare.

In addition, the distribution of the improved HDI values at the county level in China is similar to that at the county level in Shanxi, that is, the economic development level of the county shows a multi-centered, divergent, and decreasing pattern to the surrounding areas. For the city itself, the agglomeration of municipal districts with a high level of economic development in the city fully illustrates the radiation and driving role of economic development. For cities, provincial capitals have natural advantages in the allocation of resource elements, such as capital, technology, talent, and labor. Economically, as a factor resource allocation center, it is easy for provincial capitals to gather resources, and they have close cooperation with the main municipal districts of surrounding cities, which can play an obvious demonstrative and driving role.

3.4. A Study on the Spatial Distribution Pattern of County Economies in China

3.4.1. Global Spatial Autocorrelation Analysis

According to the above analysis results, the spatial distribution of improved HDI at the county level in China has a certain agglomeration trend, so we used the global spatial autocorrelation method to explore the pattern. Based on the obtained county-level improved HDI data in China, we used ArcGIS 10.8 software to calculate the global autocorrelation coefficient (Moran’s I value) of the improved HDI values of 2878 counties in China. The results are shown in

Table 3. Moran’s I index of improved HDI at the county level in China.

Global Moran’s I
Moran’s I index	0.2556
Expectation Index	–0.0003
Variance	0.000003
z-score	155.0671
p-value	0

.

In 2020, the Moran’s I index of the improved HDI at the county level in China was positive (0.2556) and passed the significance test p-value < 0.01 (p = 0), verifying that 99% of the county-level improved HDI values were not spatially random. At the same time, the z-score > 2.58 (z = 155.0671) indicated that 99% of the studies verified that the improved HDI at the county level is not randomly distributed, and the probability of random distribution is less than 1%, which is statistically significant. In order to show that the improved HDI at the county level in China has spatial agglomeration, it was necessary to explore the local spatial autocorrelation and further explore the spatial agglomeration characteristics of the improved HDI.

3.4.2. Hot Spot Analysis

We used ArcGIS 10.8 software to calculate the improved HDI values of counties in China in 2020 and obtained the local spatial correlation index (Getis–Ord Gi*) to divide the economic areas of counties in China into cold and hot spots with different degrees (Figure 5). The higher the confidence level of the hot spot, the higher the spatial correlation. The higher the cold spot confidence, the lower the spatial correlation. It can be seen in the figure that the eastern hot spot areas with high confidence are mainly distributed in coastal areas, such as the Bohai Economic Circle, the Yangtze River Delta Economic Circle, the Pearl River Delta Economic Circle, and the Leizhou Peninsula; the central areas are distributed in the Zhongyuan urban agglomeration, the southern part of Shanxi Province, and the eastern part of Hubei Province; and the western areas are distributed in the urban agglomeration on the northern slope of the Tianshan Mountains, the western part of Inner Mongolia, and the central part of Sichuan Province. All these regions have a high degree of spatial correlations. The areas with high cold-spot confidence are concentrated in the inland areas of central and western China and are mostly located in ecologically sensitive areas, inter-provincial border areas, and mountainous areas in mainland China. The cold spot area even formed the characteristic of a continuous distribution, which is concentrated in the rocky desertification area of Yunnan, Guangxi and Guizhou, the Wuling Mountains, the Hengduan Mountains, and the inter-provincial border areas in the central and southern regions. The spatial correlation of these areas is low. In 2020, there was a high degree of coupling between the improved HDI at the county level and the Gini Coefficient of nighttime lights in China, and there was a significant difference between the economically developed eastern region and the western region [38,39,40].

The analysis of the research results shows that the populations and economies of the developed areas of the eastern coastal area are highly dense. In addition to the economies of scale and economic agglomeration required by industrialization itself, this is also related to China’s accession to the World Trade Organization (WTO), undertaking the world’s manufacturing capacity, and its emergence as the “world’s factory” [41]. Thanks to the country’s policy of giving priority to the development of coastal areas after the reform and opening up, coupled with the geographical location, convenient shipping, and many ports, the coastal areas took the lead in joining the world system. Over time, the secondary and tertiary industries in the coastal areas expanded to the periphery, creating the industrial agglomeration. Most industries have formed a complete industrial chain while saving costs and improving production efficiency and economic benefits. The eastern coastal region has formed a coastal economic belt through continuous ruralization and urbanization, and each of its counties has actually become an intrinsic organic part of the urban system [42]. In addition to the coastal areas, most of the hot spots in the central and western inland areas are located in national urban agglomerations, such as the Chengdu–Chongqing urban agglomeration, Zhongyuan urban agglomeration, and the urban agglomeration on the northern slope of the Tianshan Mountains. Urban agglomerations are usually supported by national policy conditions and by strengthening inter-regional cooperation, optimizing the optimal allocation of resources and industrial structure, and dispersing the population and industrial pressure of a single large city. They also drive the construction of economic agglomeration at the county level within the region, so the economic ties in these areas are strong [43]. In the Yangtze River Delta (YRD) city clusters, the high-level core cities, connected with the surrounding cities and counties into a close group, presenting the Shanghai–Suzhou, Hangzhou–Ningbo, and Hefei clusters presenting a dual-center radiant type coexisting with a siphon type, while the Nanjing cluster is a single-center radiant type, with a synergistic and high-quality development of the YRD city clusters and the counties [44,45,46].

With the continuous growth of China’s economy, the industries in the coastal areas have gradually become saturated, the policy preferences have been reduced, and the land prices and labor costs have risen, meaning that the “backward production capacity” with a relatively low added value has gradually moved away from the core areas of the developed areas in the eastern coastal areas [41]. At the same time, the policy advantages of the Strategy for Large-Scale Development of Western China have enabled the central and western regions to adopt some of the backward production capacity that has been eliminated. However, due to the lack of talent and economic support, as well as the lack of geographical locations and topographical conditions, the industrial development of the central and western regions is relatively scattered and difficult to agglomerate, so it is impossible to form a strong economic linkage. According to the western corridor as a result of the study, there are obvious geographical characteristics of the economic links of the western city agglomeration, and it is necessary to improve the imbalance of the economic links of the city agglomeration, and fully exploit the resource advantages and location advantages [47,48]. Moreover, cold spot areas are ecologically sensitive, and topographical conditions are often more complex. The resulting inconvenience of transportation and the formulation of ecological and environmental protection policies also limit regional economic development to a certain extent.

4. Discussion

4.1. Comparison with the Existing Literature

Although most studies have used NTL to invert socio-economic development, there are few studies on inverting the improved HDI from NTL, and the objectivity of reflecting the level of socio-economic development needs to be improved [29]. As an emerging economy, China has undergone rapid economic growth following the initiation of reform and opening-up policies. The study of economic differences in China’s counties will help us gain a deeper understanding of China’s economic development at the microlevel, understand the development policies of the counties and their characteristics, and provide an economic model for high-level counties.

In this study, we constructed a multi-dimensional representative improved Human Development Index to measure the development differences of counties in China, used the nighttime light data and improved HDI in Shanxi Province as the training variables, used the data from Ningxia as the verification basis, which was different from the analysis of a single socio-economic development level using nighttime light in previous studies, and finally verified that the nighttime light data can be deduced from the improved HDI. A more comprehensive and objective level of county development is obtained, and the impact of different development conditions on the county economy of poor counties, provincial border counties, and mountainous counties is emphatically discussed, which provides reference for further research on narrowing the county gap.

4.2. Characteristics of Spatial Heterogeneity of Improved HDI in Several Specific Types of Geographical Areas

According to the list of national poverty-stricken counties published by the Poverty Alleviation Office of the State Council of China in 2014, 832 poverty-stricken counties and 2046 non-poor counties have been identified. In this study, provincial border counties and non-provincial border counties were determined according to the regulations of Cao Xiaoshu et al. (2018) [49] on provincial border counties. This study focuses on the administrative units of counties adjacent to the provincial administrative units (including provinces, autonomous regions, and municipalities directly under the central government). Provinces such as Hainan Province, Shandong Province, and Taiwan Province, which are adjacent to each other across the sea, were not considered in this study.

Multiple paired comparison analysis was employed to evaluate whether there were notable variations in the average improved HDI values among the three groups under diverse circumstances. The asterisks represent the difference between the contrasting things. The more the number, the more significant the difference. According to the results, the p-values of the improved HDI comparison data among the three groups of non-poverty-stricken counties and poverty-stricken counties, non-interprovincial border counties and inter-provincial border counties, and non-mountainous counties and mountainous counties were all less than 0.001, and the differences were significant (Figure 6).

The mean improved HDI values for the counties categorized as non-poverty-stricken and poverty-stricken are 0.2046 and 0.1108, respectively. The mean improved HDI values for the non-interprovincial border counties and inter-provincial border counties are 0.2 and 0.1381, respectively. Non-interprovincial border counties are driven by the economies of central counties or developed counties close to them and have higher improved HDI values; inter-provincial border counties have lower improved HDI values due to their distance from economic centers, backward economic development, and weaker capacity for policy implementation and self-regulation. The mean improved HDI of non-poverty-stricken counties in the non-interprovincial border counties is 0.2205, and that of poverty-stricken counties is 0.1146. The mean improved HDI of non-poverty-stricken counties in the inter-provincial border counties is 0.1574, and that of the poverty-stricken counties is 0.1062. The administrative walls and border effects of interprovincial margins on regional development limit integration and balanced development within the inter-provincial marginal zones. Therefore, even though China has provided a lot of policy assistance to poverty-stricken counties, there are still significant development differences between the poverty-stricken counties and the non-poverty-stricken counties within these regions.

According to Figure 6, counties with non-interprovincial borders account for a large proportion of the total number of counties in China, but a large proportion of these are non-poverty-stricken counties; at the same time, a large part of these counties are non-mountainous counties, which have a higher overall level of improved HDI than the rest of the counties. The inter-provincial border counties are only a small part of China’s counties, but the main poverty-stricken population is concentrated here. Meanwhile, many of the inter-provincial border counties are also mountainous counties. So, overall, the characteristics of inter-provincial borders, poverty, and mountainous counties seem to be more definitively concentrated in a certain category of counties. The development of these counties as a whole restricts the balance of China’s regional development, and only by promoting the sustainable development of these regions can the Two Centenary Goals set by China be achieved in a macro sense.

4.3. Analysis of Potential Causes of Improved HDI Spatial Heterogeneity Characteristics

Non-interprovincial border counties have always had an absolute advantage in terms of economic development due to their geographical and transport locations, which have better conditions for economic collaboration. Cooperation and development opportunities in non-poverty-stricken counties can attract more talents and labor and promote a virtuous cycle of the county’s economy. On the other hand, inter-provincial border counties have complex topographical conditions and lack funds, making it difficult to improve the transport conditions. Among all the factors, transportation and location factors play the greatest role in the development of economic differences between inter-provincial border counties and the core areas [50]. In addition, inter-provincial border regional cooperation is often affected by problems such as different administrative affiliations, ununified policies and standards, imperfect cost sharing and benefit sharing mechanisms, and imperfect cooperation mechanisms, resulting in insufficient action power of local governments [51]. The combination of factors further hinders the development of inter-provincial border counties and reinforces their marginalization. Non-poverty-stricken counties have a certain degree of independent capacity compared to poverty-stricken counties, and their production and living conditions are relatively favorable, but there is still a large gap between them and the non-poverty-stricken counties in the non-interprovincial border counties (Figure 7).

Based on whether they are mountainous counties or not, the improved HDI of mountainous counties is lower and significantly lower than that of non-mountainous counties. We mapped the dichotomous factors in terms of altitude and improved HDI (Figure 8), and it is clear that the higher and more concentrated the improved HDI in the North China Plain region is significantly related to its flat topography, where fertile land is favorable for agricultural production and more conducive to the attraction of industrial capital and enterprises. In turn, it also has indirect impacts on the development of social and public utilities. Looking at the higher-altitude regions in western China, only Urumqi on the northern slopes of the Tianshan Mountains and the Sichuan and Chongqing regions in the Sichuan Basin have high improved HDI areas. The overall altitude of these two places is lower in the west, and at the same time, the topography of the land is flat, which makes them high improved HDI areas. The overall improved HDI is lower in both the Tibetan Plateau and Inner Mongolian Plateau regions, suggesting a potential influence of topography on the improved HDI.

4.4. Limitations and Future Research Directions

Although we achieved reliable results in inverting improved HDI spatial developmental imbalances based on NTL and machine learning methods, there are still some limitations to be overcome. Firstly, the accuracy of the model results is affected by the missing data of some improved HDI indicators and the quality limitations of NTL. Secondly, although NTL data have been widely used, there are potential errors in the NTL data due to light pollution, climate and weather, human activities, etc. Third, we only applied this model to study Chinese counties, and it still has limitations when applied to the inversion and expansion of other regions globally.

In future research, we plan to develop an improved Human Development Index using a more multi-dimensional indicator system. This approach aims to provide a more equitable representation of regional development levels and address the limitations of the traditional HDI, such as enhancing the diversity of indicators and overcoming correlations among them. For instance, we propose incorporating indices like China’s social security index and carbon emissions index to highlight the development characteristics and sustainability unique to China. Additionally, when applying machine learning techniques to estimate the improved Human Development Index, we will place greater emphasis on the stability of the models and their adaptability across different regions, ensuring that the accuracy of the models is maintained when applied to other areas.

4.5. Policy Insights

Through the study, we have learned that the socio-economic development of China’s counties still has a large gap between the east and the west, but this does not mean that one side has to stop. The authorities should consider continuing to promote scientific and technological innovation in the east, continue to promote the “Western Development Strategy”, and improve the pattern of opening up to the outside world, so as to enable them to make progress together. Secondly, the socio-economic development of neighboring counties is strongly correlated, especially in economic circles and city agglomerations, where counties are more obviously driven by the economy, while inter-provincial borders, mountainous areas and ecologically sensitive areas are often neglected and difficult to develop. Based on this, our study suggests guiding the population from inter-provincial border areas, mountainous regions, and ecologically sensitive zones toward urban agglomerations and metropolitan areas, and endeavoring to improve the re-employment capacity of the mixed population, thus promoting the integrated development of urban and rural areas. In addition, for inter-provincial borders, mountainous areas, and ecologically sensitive zones, topography, transportation, and geographic location are important factors constraining their development. A possible development direction analyzed through the study is to take advantage of biological resources, agricultural products, and intangible cultural heritage to increase the value transformation of ecological resources. This can be achieved through institutional and technological innovations that unlock development potential, enabling the monetization and marketization of resources and ecological products. At the same time, it is advisable for the authorities to use regions such as the Qinghai–Tibet Plateau and ecologically sensitive zones as demonstration sites for guiding the population from predominantly agricultural activities toward non-agricultural industries, thereby enhancing the adaptive capacity of the livelihoods of the inhabitants of these special areas and continuing to promote sustainable rural revitalization in deeply impoverished areas.

5. Conclusions

In this study, we combined the improved HDI at the county level in China and NTL to conduct an inferential study of China’s counties’ economic development from a more objective perspective. The results show that the economic distribution of China’s counties presents a polycentric and decreasing law of economic hierarchy, and there is a certain agglomeration trend in the spatial distribution. Further analysis revealed that regions with high economic spatial correlations are mostly distributed in China’s national urban agglomerations, particularly in the eastern coastal areas. Regions with low correlations are predominantly located in ecologically sensitive areas, inter-provincial border areas, and mountainous areas, even forming continuous distribution characteristics. By further analyzing the possible influencing factors, we found that there are significant differences in the comparison of the improved HDI between the three groups of non-poverty-stricken counties and poverty-stricken counties, non-interprovincial border counties and interprovincial border counties, and mountainous counties and non-mountainous counties, which shows that geographic location, topographic conditions, and border effects have a greater impact on the counties’ economies. In general, this study demonstrates the correlation of economic development among the counties and regional disparities in development, and analyzes the potential causes of spatial heterogeneity of improved HDI among different regions, providing a theoretical basis for local governments to promote regional characteristic development.