Monitoring and Analysis of Population Distribution in China from 2000 to 2020 Based on Remote Sensing Data

Abstract

Accurately and precisely grasping the spatial distribution and changing trends of China’s regional population is of great significance in new urbanization, economic development, public health, disaster assessment, and ecological environmental protection. To monitor and evaluate the long-term spatiotemporal characteristics of the population distribution in China, a population monitoring estimation model was proposed. Based on remote sensing data such as nighttime light (NTL) images, land use data, and data from the fifth, sixth, and seventh censuses of China, the population spatiotemporal distribution in China from 2000 to 2020 was analyzed with a random forest algorithm. This study obtained spatial distribution maps of population density at a 1 km x 1 km resolution in 2000, 2010, and 2020. The results revealed the trend of the spatiotemporal pattern of population change from 2000 to 2020. It shows that: the accuracy assessment using the 2020 census population of townships/streets as a reference shows an R² of 0.67 and a mean relative error (MRE) of 0.44. The spatial pattern of the population in 2000 and 2010 is generally unchanged. In 2020, population agglomeration is evident in the east, with a slight increase in the proportion of the population in the west. The patterns of population agglomeration and urbanization also change over time. The population spatiotemporal distribution obtained in this study can provide a scientific reference for urban sustainable development and promote the rational allocation of urban resources.

1. Introduction

The spatial distribution and change of population is the hot spot of global sustainable development research and the center of social and economic development and ecological environment protection [1,2]. Population density is the main indicator for evaluating social development and the process of regional urbanization [3]. Fine and quantitative monitoring of population density is an important topic in current sociodemographic research. The United Nations recently released the World Urbanization Prospects report, which states that more than half of the world’s population lives in urban areas today [4]. In recognition of this characteristic of human development, the 2030 Development Agenda sets a specific goal for cities: Sustainable Development Goal (SDG) 11, which aims to “build inclusive, safe, resilient and sustainable cities and human settlements” [5,6]. Therefore, sustainable development of population density is an important element in achieving this goal. With the development of society and the economy, the temporal and spatial distribution of China’s population density has undergone significant changes. A series of changes in population have profoundly affected China’s social and economic development and natural environment. Evolution and population information play an important and fundamental role in research on climate change, urbanization, regional planning, public health, and disaster management [7,8]. Therefore, the spatiotemporal fine monitoring of populations is an important topic in the study of the sustainable development of the economy and society and life and health. Generally, the traditional methods of population monitoring are mainly obtained through population censuses and spatial interpolation. China has conducted seven national censuses in 1953, 1964, 1982, 1990, 2000, 2010, and 2020. As a commonly used reliable population source, the censuses is authoritative, systematic, and normative, providing population data based on administrative districts or census areas [9]. However, the census is very labor-intensive and time-intensive work. It has a long renewal cycle, with a census every ten years. It is difficult to meet the higher temporal resolution population data required by the current sociodemographic research [10,11]. In addition, the census is demarcated by administrative boundaries. Census data represents the entire population of an administrative area. However, the spatiotemporal distribution of the population within the administrative region is unknown, which limits the role of population data in relevant research. For example, in disaster and other related research that does not rely on administrative divisions, a high spatial resolution of the grid population data is required as the database. Therefore, although the census can obtain accurate and reliable population data, its spatial and temporal resolution makes it difficult to meet the current related research.

Additionally, in the past few decades, many related studies have explored the spatial interpolation method of gridded population density [12,13]. These methods included areal weighting [14,15], intelligent interpolation [16], and symmetric mapping [7,17]. These methods have produced many gridded population datasets, enabling population mapping of geographic areas. However, the spatial interpolation method has the advantage of being simple and easy to operate. However, its premise assumes that the population distribution within the administrative division is uniform. Such a premise ignores the actual situation of spatial heterogeneity such as, the nonuniform and inconsistent local distribution of the population, which limits its accuracy and precision.

Currently, a large number of studies use new remote sensing technology to monitor and evaluate the spatiotemporal distribution of regional populations [18]. Nighttime light (NTL) images can effectively reflect the intensity of human activities and provide more spatial details of human society. To realize the temporal monitoring of human social activities [19,20]. Based on this principle, characteristic information about socioeconomics can be extracted from NTL images. The NTL data from the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP/OLS) NTL data and Suomi National Polar-orbiting Partnership Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) NTL data are stable and persistent data sources [21,22]. The advantages of NTL data in terms of a large detection range and high temporal resolution have been widely used in the fields of economic growth [23,24,25], urbanization [26,27], poverty [28,29], carbon emissions [30,31,32], electricity consumption [33,34], natural disasters [35,36,37] and air quality assessment [38,39,40]. However, there are differences in the relationship between the NTL brightness of different land use types and the spatiotemporal distribution of the population. The NTL brightness of different land use types is not comparable. Therefore, land use data can complement NTL images in population density estimation. Recent studies have shown that the combination of NTL data and land use/land cover data can predict population distribution at the regional scale [41,42,43,44]. Tan et al. [41] used land use data, census data, and NTL data to spatially match the population of each county, simulated the spatial distribution of China’s population and its changes in 2000 and 2010, and mapped the population density in 2000 and 2010. Bagan et al. [42] used gridded land use data, census data, and satellite images of NTL to study the spatiotemporal dynamics of urban expansion in Japan. The results showed that urban expansion is related to population growth and predicted population density in the Hokkaido region. In these studies, the use of land use data still has some problems such as insufficient information extraction, resulting in low model accuracy. Therefore, further study is needed to analyze the spatiotemporal distribution of population density by different land types.

In addition, most population density mapping studies use spatially assisted data such as NTL images and land cover. The population data of administrative divisions are redistributed to the grid to form gridded population densities. Wang et al. [45] used land use/cover data and NTL data, using partial correlation analysis and geographically weighted regression to obtain a gridded population density dataset of China, which can be used to analyze the spatiotemporal pattern of population density in China. Li and Zhou [46] used NTL data and land use data to create a gridded urban population dataset with a 1 km resolution in China for 2000 and 2010. Yu et al. [47] used corrected DMSP NTL images and census data to build a population spatialization simulation model. The spatiotemporal characteristics of the population distribution from 2000 to 2010 are also revealed. Such mapping of population density requires the use of known census data as prerequisite information. The census data source also needed to be addressed for convenient and fast long-time series population fine monitoring.

Existing studies have applied abundant remote sensing data in the process of population estimation. However, this often makes the model complicated and may reduce the accuracy of population estimation. It also increases the workload of the population spatialization process. There is also a relative lack of research on population spatialization based on the latest fifth, sixth, and seventh census data of China. The census is time-consuming, laborious, and costly. There are also problems and shortcomings in the information extraction from remote sensing images for large-scale fine distribution monitoring of population density. To solve the above problems, based on remote sensing data and a random forest algorithm, this study designs and constructs a population monitoring and estimation model in China from 2000 to 2020. A regional population monitoring estimation model in China was established to generate spatiotemporal distribution maps of population density with a resolution of 1 km × 1 km in 2000, 2010, and 2020. The spatial and temporal patterns of population density changes from 2000 to 2020 were revealed and analyzed. With the rapid development of China’s economy and society, the spatiotemporal distribution of the population has changed significantly. Population monitoring can be used to analyze the fine-grained spatiotemporal variation characteristics and influence mechanisms of populations. The dataset generated in this study can also help the government and policy-makers better understand the spatiotemporal distribution of population changes. In addition, it could provide a scientific reference for the rational allocation of urban resources.

2. Study Area and Data Source

2.1. Study Area

China is the largest developing country, and by 2020, China’s population ranks first in the world. China has a vast territory and is located in eastern Asia and the west coast of the Pacific Ocean, geographically located from 73°33′E to 135°05′E and from 3°51′N to 53°33′N. There are thirty-four provincial-level administrative regions in China, including twenty-three provinces, five autonomous regions, four municipalities, and two special administrative districts.

China’s land is vast and extensive, and the country can be divided into eastern, northeastern, central, and western regions due to the differences in geographic location, natural conditions, and human and economic characteristics of each region (Figure 1).

2.2. Data Source

The experimental data in this study mainly include: NTL images, land use data, and census data from 2000, 2010, and 2020, which are listed in

Table 1. The data sources and datasets used in this study.

Datasets	Format	Resolution	Sources	Access Link
NPP-VIIRS-like nighttime light data	Grid	500 m	Harvard Dataverse	https://doi.org/10.7910/DVN/YGIVCD (accessed on 20 April 2022)
Land use data	Grid	1000 m	Resource and Environment Science and Data Center	https://www.resdc.cn/ (accessed on 20 April 2022)
Census data	Table	County level city	Census Database	http://www.stats.gov.cn/ (accessed on 15 May 2022)

.

The NTL data (Figure 2) are sourced from an extended time series (2000–2020) of 500 m spatial resolution of global NPP-VIIRS-like NTL data generated by Chen et al. [48]. Chen et al. built an extended time series (2000–2020) of NPP-VIIRS-like NTL data through a new cross-sensor calibration from DMSP-OLS NTL data (2000–2012) and a composition of monthly NPP-VIIRS NTL data (2013–2020). Chen et al. first developed a modified auto-encoder model. Secondly, the architecture of the auto-encoder model with CNN was designed, and the cross-sensor calibration model was trained. Thirdly, an extended time series of the NPP-VIIRS-like NTL dataset was generated using the trained model by inputting the 2000–2012 EANTLI data and appending the composited NPP-VIIRS NTL data with postprocessing of the data. Finally, a comprehensive accuracy evaluation was conducted. Compared to the composited NPP-VIIRS NTL data in 2012, Chen et al. extended NPP-VIIRS-like NTL data show good accuracy globally at the pixel (R²: 0.87, RMSE: 2.96) and city (R²: 0.95, RMSE:3024.62) levels. At the regional scale, all countries show an acceptable accuracy. The R² ranges from 0.70 to 0.86, and the RMSE is lower than 6 ${\mathrm{nWcm}}^{-2}{\mathrm{sr}}^{-1}$. The NPP-VIIRS-like NTL data have an excellent spatial pattern and temporal consistency, which are similar to the composited NPP-VIIRS NTL data. NTL can reflect the level of economic development of the region and is widely transported to the inversion of socioeconomic parameters.

Land use data at a 1 km spatial resolution for 2000, 2010, and 2020 were obtained from the Resource and Environment Science and Data Center (https://www.resdc.cn/) (accessed on 20 April 2022). These images are mainly based on US Landsat TM images and Landsat 8 remote sensing images, and are generated through artificial visual interpretation. Land use and land cover data provide important information about parcels, including class, area, and spatial distribution [49], which involves four categories for human habitation: urban land, rural land, industrial transportation land, and arable land. Urban land refers to the built-up areas in large, medium, and small cities, counties, and towns. Rural land refers to rural settlements independent of towns. Industrial transportation land refers to the land used for factories and mines, large industrial areas, oil fields, salt fields, quarries, and other lands, as well as traffic roads, airports, and special land. Arable land is the area on which crops are grown.

Census data for 2000 and 2010 were obtained from the databases of the Fifth and Sixth Censuses. The 2020 census data were obtained from the Seventh National Census Bulletin at the county level (no data for Xinjiang in 2020). The census data used in this study are census-defined resident population data.

The latest administrative area boundaries in 2020 were selected as the basis to solve the problem of administrative boundaries and name changes of districts and counties in 2000 and 2010. We achieved consistency in the administrative area boundaries of districts and counties for these three census years. The population census data were adjusted with their inconsistent administrative area boundaries.

3. Methods

3.1. Impact Factor Calculation and Selection

NTL data are closely related to human activities and can reflect population distribution to a large extent. Many researchers have used DMSP-OLS and NPP-VIIRS NTL images to evaluate population density and population distribution [18,50,51,52]. Sun et al. [50] used DMSP/OLS NTL data to simulate urban and rural population densities. You et al. [51] analyzed the spatiotemporal evolution of the population in Northeast China based on linear regression analysis and demographic data of NPP-VIIRS NTL images. The influencing factors of population distribution are complex and diverse. Existing research results show that NTL images and land use data can effectively estimate the population of the monitoring area [41,43]. We extracted the spatial information that can reflect the population distribution phenomenon as much as possible and addressed the needs of the subsequent downscaled remote sensing inversion. In this study, the total value of NTL and the total number of luminous grids in the sample counties were selected as the feature factors of NTL images. Among them, the former reflects the overall brightness level of regional lights, and the latter can reflect the approximate area of human activities in the region. Additionally, these two characteristic factors are the statistical values of the regional total, corresponding to the scale of the regional total population.

Land use data can reflect the current use direction and development degree of the surface, describe the composition of the surface landscape, and reflect the intensity of surface use. There are differences in the population carried by the areas of different land types. If only artificial surface area is used as the characteristic factor of land use data, it is difficult to reflect the real situation of population distribution. Related studies have shown that the ratio of woodland and grassland area has a negative impact on population density, which leads to negative values during model estimation [41]. In contrast, the introduction of woodland and grassland area ratios has no significant effect on the accuracy of model estimation [41]. Therefore, this study selected the urban land area, rural land area, industry-traffic land area, and cultivated land area in the sample areas and counties as characteristic factors of land use data (Figure 3). These four types of land use are all man-made landscapes, which can effectively reflect the current situation of human activities and population distribution. In addition, the four characteristic factors are also the statistical values of the regional total, which can meet the needs of the subsequent downscaling estimation of the model.

Therefore, the total value of NTL, total number of luminous grids, area of urban land, area of rural land, area of industry-traffic land, and area of cultivated land were finally selected as the characteristic factors in this study.

3.2. Population Monitoring Estimation Model Construction

In recent years, the method of establishing machine learning models based on multisource geospatial data to study population distribution has become very popular. Machine learning enables the estimation of continuously distributed population data from a discrete perspective. Machine learning regression methods can analyze and extract key feature information in known datasets, inspire complex correlations between search data, and repeatedly train and verify through a large number of samples for accurate simulation and prediction. One of the more mature and commonly used models for population estimation is the random forest model. The random forest model is a classic machine learning regression method based on decision trees, which has a fast execution speed and good model performance and is mostly used for classification, regression, and clustering in machine learning [53]. It is based on the idea of ensemble learning. Based on the bagging algorithm, the data obtained by random sampling input many weak learners, and vote to obtain the final output result. The random and ensemble features of the random forest model make the result of the overall model have high accuracy and generalization performance, but also good stability.

In this study, the random forest model was selected to construct and train the estimation model of China’s regional population distribution, using the abovementioned six characteristic factors, including the total value of NTL, the total number of luminous grids, area of urban land, area of rural land, area of industrial and mining land, and area of cultivated land as explanatory variables. Using the total population of the sample districts and counties as predictor variables, the model was trained several times on the characteristic factor data and the census data of the sample districts and counties in 2000, 2010, and 2020. The 5-fold cross-validation method was used to evaluate the model prediction effect. During the training process, the model parameters with the best regression effect are selected to determine the final random forest population estimation model. According to the model training effect, the minimum leaf size was eight, and the number of learners was thirty as model parameters.

3.3. Population Estimation Model Revision

Based on the constructed population estimation model, this study extracted six characteristic factors and obtained the results of population spatialization. However, due to the light overflow phenomenon in the NTL image, light also exists in the nonartificial surface area. As a result, there is population distribution on the nonartificial surface in the estimation result, but there is generally no population distribution on the nonartificial surface in reality. Therefore, in this study, the population number of the nonartificial surface area in the estimated population distribution result was set as zero. In addition, the total population of districts and counties obtained from the population estimation model may differ from the actual total population of districts and counties. Therefore, it is necessary to make statistical corrections to the results of the population estimation model to make the estimated population results consistent with the actual population at the district and county scales. Such population spatial distribution correction results can more accurately reflect the real population spatial distribution. The specific population monitoring estimate revision formula is as follows:

$$a = \frac{P_{\mathit{ja}}}{P_{\mathit{je}}}$$

(1)

$${\mathit{POP}}_{\mathit{ja}} = a \times {\mathit{POP}}_{\mathit{je}}$$

(2)

In the formula, $a$ is the correction factor, ${\mathit{P}}_{\mathit{ja}}$ is the actual total population of district j, ${\mathit{P}}_{\mathit{je}}$ is the model estimated total population of district j, ${\mathit{POP}}_{\mathit{ja}}$ is the corrected population result of grid j, and ${\mathit{POP}}_{\mathit{je}}$ is the model estimated population of grid j.

3.4. Precision Inspection and Verification

Accuracy evaluation is the verification that reflects the accuracy of the spatialized model and is an important step in building the model [45]. We take the 2020 census data of 800 townships/streets in Beijing, Shanghai, Guangzhou, Shenzhen, Anhui Province, Shaanxi Province, Gansu Province, Guizhou Province, Jilin Province, and Hubei Province as examples to evaluate and verify the accuracy of regional population monitoring in China in this study. We mainly choose the mean relative error (MRE) to calculate the accuracy.

$$\mathit{MRE} = \frac{1}{n}{\sum }_{i = 1}^n|\frac{y_{\mathit{act},i} - y_{\mathit{est},i}}{y_{\mathit{act},i}}|$$

(3)

In the formula, $y_{\mathit{act},i}$ is the census population of the ith region, and $y_{\mathit{est},i}$ is the estimated population of the ith region.

4. Results

4.1. The Spatiotemporal Pattern of China’s Regional Population Density Distribution from 2000 to 2020

Based on the above modeling methods, using NTL data, land use data, and census data, we obtained population density maps in mainland China in 2000, 2010, and 2020 with a spatial resolution of 1 km. In recent decades, China has experienced rapid socioeconomic development and accelerated urbanization. However, there is obvious spatial variability in China’s economic development and significant regional development imbalance, which has led to significant differences in population growth patterns and the spatial and temporal distribution of the population in eastern, central, western, and northeastern China. The population density maps of China in 2000, 2010, and 2020 (Figure 4) show that the spatial distribution patterns of China’s population in these three years are similar. The spatial distribution pattern of China’s population did not change significantly from 2000 to 2010. In 2020, the population of eastern coastal cities and some developed cities increased significantly.

From the population density map of mainland China in 2000, 2010, and 2020, it can be seen that there are obvious differences in population distribution in different regions, mainly due to differences in regional economic growth rates. Economically developed cities have relatively complete infrastructure such as medical care, education, transportation, and entertainment, making life more convenient. In addition, a developed economy means that cities have more employment opportunities and can attract more people to settle down. Therefore, the population density is higher in the more economically developed eastern coastal areas of China, especially in highly developed Chinese cities such as Beijing, Shanghai, Guangzhou, and Shenzhen. In addition, the provincial capitals and port cities are also hotspots for population distribution, most of which have a population density greater than 3000 people/km².

Relatively speaking, the population density in western China as a whole is very low except for provinces and cities such as Sichuan Province and Chongqing City, where high population densities exist. In particular, the provinces of Tibet, Xinjiang, and Qinghai in China are extremely sparsely populated in the region, with obvious characteristics of extensive land and sparse people. Most of the population densities in the west are below 50 people/km². The main reason for the sparse population in western China is that it is difficult for natural conditions to meet the demands of rapid economic development. The population loss is serious, which in turn makes the local labor force insufficient to promote economic development, resulting in a scarcity of jobs. The population will in turn migrate to economically developed areas, forming a vicious cycle of population loss [54,55].

The central and northeastern regions of China connect the economically developed east and the less economically developed west, so the population density of the central and northeastern regions is between the east and the west. The proximity of some provinces in central China to economically developed eastern provinces and cities, coupled with an excellent economic development environment, makes these provinces significantly more densely populated than surrounding regions, such as Hubei, Hunan, Henan, and Hebei provinces. The economic development of the Northeast region has stagnated, and the population has lost a lot, making the population density of the Northeast region significantly lower than that of the eastern region. The spatial difference in population density in China is mainly due to unbalanced economic development, topography, and landscape.

Based on the 1 km population density distribution, the average population density for urban level (Figure 5) and provincial-level regions (Figure 6) in China were obtained. The results show that the distribution of urban population density in China in 2000, 2010, and 2020 is similar, with an overall pattern of being high in the southeast and being low in the northwest. The population density of coastal cities in southeast China is significantly higher than that of other regions. Among them, the cities in the Pearl River Delta and Yangtze River Delta urban agglomerations have a relatively high concentration of high population density cities. In addition, the population density of cities in central China is significantly higher than that in western and northeastern China. The average population density distribution pattern at the provincial level in China in 2000, 2010, and 2020 is similar to that at the city level. The overall pattern is high in the southeast and low in the northwest.

4.2. China’s Regional Population Growth Trend from 2000 to 2020

Between 2000 and 2010, China’s population increased by 73.9 million people, and between 2010 and 2020, China’s population increased by 72.06 million people. Figure 7 shows the population growth rate of each province in each decade, and the provinces with growth rates are basically in the eastern region. From 2000 to 2010, Beijing and Shanghai had the largest population growth, with growth rates of 0.42 and 0.38, respectively. From 2010 to 2020, Guangdong and Tibet Provinces had the largest population growth, with growth rates of only approximately 0.2. The population growth in 2010–2020 was generally less than the population growth in 2000–2010, and the number of provinces with negative population growth increased. Harbin Province has the highest negative population growth with a negative growth rate of 0.17.

Based on the population distribution results with a 1 km spatial resolution, we calculated the population change results in China from 2000 to 2010 and from 2010 to 2020 with 1 km spatial resolution (Figure 8). This result showed that population growth mainly occurred in the Yangtze River Delta, Pearl River Delta and Beijing-Tianjin-Hebei regions. With the rapid development of China’s economy and the continuous advancement of urbanization, China’s overall population density has grown rapidly. In particular, highly developed cities such as Beijing, Shanghai, Guangzhou, Shenzhen, and Chongqing are attracting larger populations.

From 2000 to 2020, the trend of population density changes in western China was similar. Only some regions change, such as Sichuan, Chongqing, and Yunnan Provinces, and cities. However, there are more regions with negative population growth. Other western regions such as Xizang and Qinghai provinces did not show significant population changes. The population density in some regions of Northeast China had a certain positive growth from 2000 to 2010, but it was more of a negative growth trend. In 2010–2020, the number of negative growth areas increased, and the overall population density showed significant negative growth. From 2000 to 2020, there was a clear growth difference in central China. Provinces such as Hunan and Hubei have more negative growth, while Henan and Hebei have more positive growth. From 2000 to 2020, the population density of the three major urban agglomerations in the Yangtze River Delta, the Pearl River Delta, and the Beijing-Tianjin-Hebei region in eastern China increased rapidly, but there was still significant negative growth in areas such as Jiangsu Province.

Population changes show the imbalance of regional development, and there are obvious local population differences between cities. China’s population changes over the past 20 years can reflect the development of China’s urban economy and rapid urban expansion. The population of many economically developed cities has increased by more than 1000 people/km². Therefore, demographic changes can be applied to the analysis of urbanization or urban issues, bringing some new insights into urban development.

5. Discussion

5.1. China’s Regional Population Growth Trend from 2000 to 2020

Since the reform and opening up, China has experienced rapid urbanization, developed very rapidly, and experienced the largest urban-rural population migration in human history [56]. Population mobility exacerbates regional disparities, thereby increasing the complexity of population distribution. Population mobility and population distribution will change people’s way of life and production, and affect social and economic development. Some studies have shown that economic development and population distribution are highly correlated and that the two interact with each other. Therefore, unbalanced regional economic development also leads to unbalanced regional population distribution, which also poses a huge challenge for the Chinese government to balance the development between regions.

Rapid economic and social development is accompanied by rapid population growth, while explosive population growth also drives rapid economic and social development. From 2000 to 2010, the spatial distribution pattern of China’s population density remained basically unchanged, and the population density of the eastern coastal cities increased slightly. From 2010 to 2020, the spatial distribution of China’s population showed a trend of gradually decreasing population from economically developed cities, provincial capital cities, and port cities to peripheral areas. Among them, the eastern region has the highest population concentration, and the population density of the western regions, such as Tibet, and Inner Mongolia remains unchanged. High-value population density areas in China are more often located in economically developed regions, such as the Yangtze River Delta, the Pearl River Delta, and the Beijing-Tianjin-Hebei urban agglomeration. The economic development model of the vast majority of Chinese provinces is a single-core or dual-core development model. Such a development model has led to an excessive concentration of public resources and capital in the core cities, and a large number of jobs and public services generated by them has attracted a large number of people, resulting in a significant siphoning effect of the core cities on the neighboring cities. As a result, core cities may slow down the development of neighboring cities while promoting their own economic development, resulting in a continuous outflow of population from less economically developed cities. The change in regional population reflects the uneven regional economic development, and the distribution of population density corresponds to the level of urban economic development.

5.2. Population Monitoring Estimation Model Validation Results

In this study, a stochastic forest population estimation model was constructed based on remote sensing data. The model results were evaluated using a fivefold cross-validation, which means that 20% of the sample data from districts and counties were randomly selected for model validation. The verification results are shown in Figure 9. The R² in 2000, 2010, and 2020 were 0.63, 0.70, and 0.78, respectively. This method proves the use of NTL data, land use data, and census data to simulate and estimate the population space at the county level in China. The validity of the distribution allows for fine-scale monitoring of the spatial distribution of the population.

In addition, this study validated the accuracy of the simulated population spatialization results with townships/streets census data. Since the seventh census data at township/street level are not fully released at present, we use the census data of 750 townships/streets in Beijing, Shanghai, Guangzhou, Shenzhen, Anhui, Shaanxi, Gansu, Guizhou, Jilin, and Hubei to validate the accuracy of the spatialized results in 2020. We also compared with two population grid datasets (GPW and Worldpop) and the validation results are shown in Figure 10. The R² for this study was 0.67, the mean relative error (MRE) was 0.44 and the p value was 0.00, while the R² for the GPW dataset was 0.55 and the MRE was 0.99, and the R² for the Worldpop dataset was 0.60 and the MRE was 0.59. The accuracy and mean relative error of the predicted values in this study are better than the results of GPW and Worldpop datasets. The simulated population spatial distribution results of this study are highly correlated with the census results.

5.3. Challenges in China’s Regional Population Monitoring and Estimation Based on Remote Sensing Data

In recent years, the rapid development of remote sensing technology has provided us with higher and finer resolution data, such as NPP-VIIRS and Luojia-1 NTL images, and a large number of studies have used NTL images for population estimation. However, these remote-sensing images still have some limitations. With economic and social development, a large number of light sources discharge from lamps and LEDs, and the difference between the spectra of these two sources has produced a great deal of uncertainty in NTL images. The spatial resolution of the nighttime light images is also low, only 500 m. In addition, there are blooming effects in the nighttime light images. In the future, we will explore higher-resolution NTL images and strive to improve modeling methods to improve the accuracy of population spatialization. We will consider using methods such as spatial autocorrelation models to solve the blooming effects problem of nighttime light images [57]. Land use data also have some limitations such as low spatial resolution. It is difficult to achieve accurate land class information extraction from 1 km resolution images. In the future, we will consider using high-resolution remote sensing images to classify the land into more accurate and fine land classes.

In this study, we use the random forest model as the population density estimation model. Compared with the models used in previous studies such as OLS and spatio-temporal geographically weighted regression models, the random forest model can better solve the problem of negative population in the process of population spatialization. In addition, the random forest model is also able to better solve the problems of overfitting and multicollinearity of the explanatory factors [58]. Moreover, the random forest population density estimation model constructed in this paper has been verified to have high estimation accuracy. Although random forest is a well-performing and generalizable algorithm, it may not be able to solve the problem of spatial heterogeneity [59]. China is a vast country and the population distribution may be spatially heterogeneous, and the random forest model is less able to infer the potential spatial heterogeneity among them. There may be some bias in the population estimation at a large regional scale, the population linkage among regions is not reflected, and the overestimation and underestimation of urban and rural populations should be further explored. In the future, we will add multisource data to explore the spatiotemporal distribution of China’s precise population, solve the problem of overestimation and underestimation of urban and rural populations, and improve the accuracy of population estimates.

6. Conclusions

In this study, we developed an estimation model for regional population monitoring in China based on a random forest algorithm using nighttime lighting data, land use data, and population census data. Based on this model, we realized 1 km resolution population density mapping in China in 2000, 2010, and 2020. We revealed and analyze the spatiotemporal distribution of the regional population and its change characteristics pattern over 20 years. The results showed that:

(1) The accuracy assessment using the 2020 census population of townships/streets as a reference shows an R² of 0.67 and an MRE of 0.44. Comparing the population spatialization results simulated in this study with the GPW and Worldpop grid population datasets, the accuracy is better than both of them. It shows that the method of this study can produce reasonable accuracy.

(2) The spatial pattern of the population in 2000 and 2010 remained unchanged overall. In 2020, the phenomenon of population agglomeration was obvious, and the population growth in the eastern region was the most obvious. The proportion of the population in the west increased slightly. The patterns of population agglomeration and urbanization also changed over time.

(3) The population distribution showed a multi-center spatial pattern, mainly concentrated in the economically developed cities, provincial capitals, and port cities, and gradually decreased in the surrounding areas. The distribution of population density is compatible with the level of urban economic development.

Based on remote sensing data, this study constructs a stochastic forest population monitoring estimation model for Chinese regions, which can provide an important reference and technical support for the fine monitoring of regional spatiotemporal population distribution, population planning and management, and sustainable economic and social development.