Identification of densely populated-informal settlements and their role in Chinese urban sustainability assessment

ABSTRACT

China’s National Plan on Implementation of the 2030 Agenda for Sustainable Development identified the upgrading of urban shantytowns, urban villages, and dilapidated houses as an important initiative to implement the Sustainable Development Goal (SDG) 11.1. However, informal housing being used as temporary housing by low-income families (especially in megacities) has resulted in informal settlements that are usually densely populated, dilapidated, and disorganized. Identifying targets based on deep learning and single very high-resolution images remains a challenging task. Here, we used multi-source geographic information data and machine-learning methods to identify and compare the distribution of densely populated‒informal settlements (DPISs) and measured population ratios within the urban areas of three Chinese first-tier cities: Beijing, Shanghai, and Guangzhou. Our results indicate that DPISs occupy 1.98%, 0.67%, and 3.95% of the total area in these study regions, with population densities of 42,800 people/km², 7,100 people/km², and 20,800 people/km², respectively. Further, significant variability existed in the distribution of DPISs, population ratios, and composition of the study areas. DPISs reflected the rapid urbanization process in China and the planning of the city, which can be used as an indicator of sustainable urban development in China.

KEY POLICY HIGHLIGHTS

• Localizing SDGs indicators to study areas and incorporating regional variations and specific characteristics enhanced the understanding of their role in regional sustainable development.

• DPISs were introduced in this study; the similarities and differences were made between DPIS and SDG indicators.

• Contrasting distribution patterns of DPISs in Beijing, Shanghai, and Guangzhou revealed distinct regional disparities and urban development dynamics.

• Integration of multi-source remote sensing data enabled comprehensive analysis of complex features and spatial patterns of DPISs.

• Exploration of the intricate relationship between DPISs and urbanization unveiled challenges and potential opportunities for inclusive urban development.

KEYWORDS:

• SDG 11.1
• multi-source geographic information data
• machine learning
• densely populated-informal settlements
• localization indicators

Introduction

Urbanization is a hallmark of the Anthropocene (Elmqvist et al. 2021) and accelerates the process of regional interactions across global scales and domains (McPhearson et al. 2021). In recent decades, human migration has become a major driver of urbanization (Xiao et al. 2022) and has improved urban economic growth (Chen, Liu, and Lu 2018). However, urbanization has progressed far ahead of economic growth in some urban areas (such as Beijing and Guangzhou). Insufficient and underdeveloped infrastructure cannot supply the basic services necessary to satisfy the diverse needs of people. In particular, one of the major consequences of rapid urbanization in developing countries of Asia and Africa is an increase in informal settlements such as slums (Stokes and Seto 2019).

The year 2015 marked the adoption of the Sustainable Development Goals (SDGs) by the United Nations, which consists of 17 goals aimed at achieving sustainable development worldwide. Goal 11.1 states that slums should be upgraded to ensure access to adequate, safe, and affordable housing and basic services by 2030 for all using the “proportion of urban population living in slums, informal settlements, or inadequate housing” as a key indicator of urban sustainability (United Nations 2019). The global urban slum-dwelling population exceeded 1 billion in 2018; this accounts for 24% of the total urban population (United Nations 2021). Overcrowding and poor living conditions in slums are conducive to the spread of diseases, especially during the coronavirus disease 2019 (COVID-19) pandemic (Li et al. 2021; Zhang et al. 2022). The vulnerability of urban areas (especially slums) in many countries is evident, as dense populations and inadequate social distancing make them extremely susceptible to the spread of epidemics (Beevers et al. 2022; Han et al. 2020; Wang et al. 2022). It is necessary to determine the size of slums in urban areas and quantify the proportion of people living in them to guide the development of appropriate policies and programs to ensure adequate housing and slum upgrading for all, so that urban planning can be adjusted to mitigate the risk of exposure to major infectious diseases.

The global sustainable development indicator framework developed for the SDGs should be adapted to the specific situation of each country when used to evaluate national and regional sustainable development (Wang et al. 2020). Urban slums are a microcosm of poverty that is measured differently by various countries and regions of the world (Ezeh et al. 2017). Slums refer to informal settlements in developing countries that are characterized by overcrowding, unsafe living conditions, and lack of access to clean water, electricity, sanitation, and other basic services.

China is a developing country with a population of 1.4 billion, whose urbanization rate rapidly rose from 19.4% to 52.6% between 1980 and 2012 (Yang 2013). In the early days, many migrant workers that moved to the city could not afford housing. Therefore, they built temporary shelters on wastelands, roadsides, riversides and vacant lots around factories called shantytowns (Li, Kleinhans, and van Ham 2018). In addition, most of the arable land in the countryside was expropriated owing to urban expansion, and farmers still lived in their original villages after being converted into residents. The villagers’ houses were surrounded by urban buildings, and the corresponding organization and social relations were extended, forming new urban communities with the characteristics of village communities called urban villages. Most of these areas were composed of highly dense informal housing that attracted many low-income migrant workers owing to the relatively low rent (Huang, Dijst, and van Weesep 2018; Zheng et al. 2009). In September 2016, “China’s National Plan on Implementation of the 2030 Agenda for Sustainable Development” was released. It identified the upgrading of urban shantytowns, urban villages, and dilapidated houses as an important initiative to implement SDG 11.1.

Urban renewal was implemented in numerous cities in recent years as a response to disordered urban sprawl and the decay of earlier old urban areas. Many of these old informal houses were demolished or rebuilt. However, the renewal of numerous old settlements is still in the preliminary planning stage owing to a large number of subjects and long project cycles involved. In particular, informal housing (primarily in shantytowns or urban villages) is still used as temporary housing by low-income families in some megacities, resulting in informal settlements that are usually densely populated, dilapidated, and disorganized with low community greenery that poses greater safety risks (Li and Wu 2013; Liu et al. 2010; Wu 2008). Therefore, it is necessary to identify and quantify the proportion of the population living in such settlements and explore its role in measuring sustainable urban development in China.

Previous studies have used machine learning to obtain details of urban scenes from large-scale remote sensing images to enhance the precision and automation of the identification of slums and informal settlements (Ajami et al. 2019; Huang, Davis, and Townshend 2002; Xu et al. 2022). Very high-resolution (VHR) remote sensing images with a sub-meter pixel resolution contain a large amount of spectral, structural, and textural information. Such images provide information on the composition of urban land use, detailed urban landscape, and building information (Li et al. 2016; Pacifici, Chini, and Emery 2009; Sohl et al. 2019). Deep learning (DL) is emerging as the model of choice in many aspects of VHR applications (You et al. 2018), such as scene classification (Liu et al. 2018), object detection (Cheng and Han 2016), image retrieval (Gordo et al. 2017), and data fusion (Schmitt and Zhu 2016). Urban scene classification aims to automatically assign a semantic label to each blocked image; it is an active research topic to identify different types of complex urban scenes (Huang, Chen, and Gong 2018). However, different urban scenes can contain the same types of objects or share similar spatial arrangements. For instance, historical residential areas, apartment blocks, and low-income residential neighborhoods may contain low-rise houses, even though they are different scene types. Therefore, identifying targets based on DL and single VHR images is a challenging task.

Multi-source geographic information data with macro-dynamic and multi-scale monitoring capability is an important tool for the evaluation and achievement of target SDG indicators (Jochem, Bird, and Tatem 2018; Liu et al. 2017; Zhao and Zhang 2018). Geospatial big data (such as nighttime light satellite imagery, vegetation cover, land cover data, and location-based data) can simulate various features of urban landscapes (Yu et al. 2014), environment (Bino et al. 2008), infrastructure (Cole et al. 2017), and population density (Pulselli et al. 2008). This adds potential sources to improve the accuracy of urban scene identification. Currently, two types of nighttime light satellite data are used to characterize urban spatial patterns, and structures and estimate socioeconomic SDG indicators: the Defense Meteorological Satellite Program Operational Linescan System and the Suomi National Polar-Orbiting Partnership Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) (Li et al. 2019; Ma et al. 2012; Small, Pozzi, and Elvidge 2005). Newly released data from the Luojia1–01 satellite produced by Wuhan University have a higher spatial resolution of 130 m and show potential for more details of inner-city structure (Li et al. 2018). Vegetation cover and land cover data are mainly derived from multi-spectral satellite imagery, such as Landsat-8 multi-spectral imagery, and are used to quantify urban green space (Li, Saphores, and Gillespie 2015), spatial extent of urban area (López et al. 2001), and artificial impervious area (Gong et al. 2020).

Our study aimed to accurately identify and map densely populated-informal settlements (DPISs) within urban areas using multi-source geographic information data and machine-learning techniques. This objective fulfills an important gap in knowledge regarding the spatial patterns and characteristics of DPISs in China, contributing to the field of urban studies and sustainable development. Moreover, insights into the social and environmental challenges associated with rapid urbanization are gained by quantifying population ratios in DPISs. This knowledge facilitates the development of targeted urban planning and policy interventions, fostering improvements in living conditions and access to basic services for the population residing in DPISs. The ultimate goal is to contribute to the creation of more sustainable and inclusive urban environments in China, with the potential for broader applications in similar contexts worldwide.

Materials and methods

Study area

The main urban areas of three cities in China, Beijing, Shanghai, and Guangzhou, were chosen as the study area (Figure 1). The study conducted in Beijing selected the specific area enclosed by the Fifth Ring Road. This area covers approximately 670 km² and represents 4.1% of the entire urban region. The study area for the research conducted in Shanghai was limited to the outer ring, covering approximately 10.4% or 660 km² of the entire urban area. The study area in Guangzhou was bound by the bypass highway, covering 770 km² and approximately 10.4% of urban regions.

Figure 1. Specific study areas in three Chinese cities: (a) Beijing, (b) Shanghai, and (c) Guangzhou.

Notes: Urban and rural land, industrial and mining land, residential land area data were extracted from the 2020 China Land-Use Remote Sensing Monitoring data, published by the Resource and Environmental Science and Data Center of the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences (https://www.resdc.cn/data.aspx?DATAID=335), with a spatial resolution of 30 m.

Rapid urban development has led to the formation of slums, urban villages, and informal settlements. The availability of good employment opportunities has attracted a large number of migrants, making the impact of DPISs more typical in megacities. In 2020, the resident population in the study areas of Beijing, Shanghai, and Guangzhou was 9.41 million, 6.68 million, and 6.34 million, respectively. This accounted for 43%, 27%, and 34% of the total population of each city, respectively.

Datasets

The datasets used in this study (including multi-source remote sensing imagery and auxiliary geospatial data, were all obtained from publicly available data with free access. Multi-source remote sensing imageries were obtained from Gaofen-2 satellite images, Luojia-1 satellite nighttime light images, and Sentinel-2 images (Table 1). The multi-spectral and panchromatic band fusion of the Gaofen-2 satellite images (with a spatial resolution of 0.8 m) were used as scene information for deep learning. Images in the experiment were segmented into several images of 128 × 128-pixel size. Luojia-1 satellite images were acquired free of charge from the Hubei Provincial Data and Application Center, with a spatial resolution of 130 m. Sentinel-2 images with 10 m spatial resolution were freely obtained from the European Space Agency website.

Table 1. Experimental data sets.

No.	Dataset	Year	Spatial resolution	URL
1	Gaofen-2	2019	0.8 m	http://www.cresda.com/CN/index.shtml
2	Luojia-1	2018	130 m	http://59.175.109.173:8888/
3	Sentinel-2	2019	10 m	https://scihub.copernicus.eu/
4	RTUD	2019	1,000 m	https://heat.qq.com/index.php
5	Road network data	2019	—	http://www.openstreetmap.org/
6	WorldPop dataset	2019	100 m	https://www.worldpop.org/
7	China Land-Use Remote Sensing Monitoring Data	2019	30 m	https://www.resdc.cn/-data.aspx?DATAID=335

Available geospatial data included real-time Tencent user densities (RTUD), road network data, and the WorldPop dataset. Real-time Tencent user densities can reflect the population density of the region. They were obtained from the Tencent travel spatial data platform interface. The coverage time was 21:00‒24:00 on 26 August 2019. The RTUD records the locations of smartphone users of Tencent applications, such as Tencent Mobile App QQ (a messenger-like software), WeChat (a mobile chat software), Soso Maps (a web mapping services and navigation software), and some other mobile applications that provide LBS services (Yao et al. 2017). Road network data were obtained freely from an open-source map called OpenStreetMap and overlaid with land use data to obtain the neighborhood distribution of urban, rural, industrial, mining, and residential land uses in the study area. The WorldPop dataset is available for free on the WorldPop Project website. The 100-m spatial resolution mapping was created as part of the Global Population Distribution Mapping Initiative, which was spearheaded by the Geodata Institute at the University of Southampton in the UK.

In addition, the 2020 China Land-Use Remote Sensing Monitoring Data (released by the Resource and Environmental Science and Data Center, Institute of Geographical Sciences and Resources, Chinese Academy of Sciences) provided data on land use in various contexts such as urban, rural, industrial, mining, and residential areas.

Paper structure

The article is structured into three main sections, each focusing on distinct aspects of the research (Figure 2).

Figure 2. Paper structure.

The first section introduces the study’s methodology and outlines the process of DPIS extraction. This involves the use of multi-source geographic data and machine-learning techniques, specifically the combined utilization of CNN and RF models. These methods are applied to effectively identify and extract DPISs across the examined urban areas.

The second section extensively analyzes the characteristics of DPISs, encompassing factors like population density and land use distribution. This scrutiny offers a comprehensive grasp of the dynamics and implications linked with DPISs within urban settings. The exploration encompasses DPIS distribution and population ratios in Beijing, Shanghai, and Guangzhou. Additionally, it delves into the correlation between shifts in urban land use and DPISs, unraveling the intricate interplay between urbanization and these informal settlements.

The third section centers on a comparative assessment between DPISs and SDG 11.1. This segment explores how DPISs align with the objectives outlined in SDG 11.1, which pertains to the upgrading of urban shantytowns and informal housing. By juxtaposing DPISs against this specific sustainability goal, the research underscores the relevance and potential impact of these settlements on achieving sustainable urban development within the Chinese context.

In conclusion, the article systematically progresses from the methodology of DPIS extraction, through an in-depth analysis of their characteristics, and culminates in an evaluation of their alignment with SDG 11.1. The overall narrative highlights the role of DPISs in the pursuit of sustainable urban development in Chinese cities, showcasing their significance within the broader framework of urban planning and policy implementation.

Methodology

Comparison of models

Naive Bayes (NB), Support Vector Machines (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), and VGG16 and VGG19 models from Convolutional Neural Networks (CNN) were compared to select appropriate methods for accurate extraction of DPISs.

The NB algorithm shows robustness and performs well when the dataset is independent, but encounters issues with correlated input variables. An accuracy of approximately 77% was achieved by the Gaussian NB algorithm with maximum likelihood estimation when DPISs and formal settlements (FSs) were used as input variables. The SVM model demonstrates strong capabilities in addressing non-linear separable problems, but it encounters challenges in training and multi-class classification. Moreover, the performance of SVM deteriorates when dealing with large training sample sizes. A total of 300 images (256 × 256 pixel size) were selected for each class, and the accuracy of the SVM model remained below 50%. The RF model demonstrates superior robustness and accuracy compared with the preceding two shallow-learning algorithms. The RF model can achieve a training accuracy of over 85% by utilizing high-resolution Google Earth image data despite the potential issue of overfitting in the presence of significant noise.

The KNN model has the advantage of simplicity, interpretability, and ease of implementation, does not require parameter estimation, and makes no assumptions about the data. This results in high accuracy and the ability to ignore outliers. However, it has a high computational cost, especially with a large number of features. The prediction accuracy is affected when the sample distribution is imbalanced. The model accuracy using unsupervised algorithms was approximately 79%. In contrast, CNN offers a concise structure: VGG16 consists of 16 hidden layers (13 convolutional layers and 3 fully connected layers), whereas VGG19 comprises 19 hidden layers (16 convolutional layers and 3 fully connected layers). The accuracy of VGG16 and VGG19 was 85% and 50%, respectively. The CNN models have high resource consumption, a large parameter size, and increased memory usage.

The RF and CNN (VGG16) approaches showed the most favorable performance in the overall comparison of the six machine-learning methods (Figure 3). Consequently, the CNN + RF model architecture was selected for subsequent study.

Figure 3. Accuracy of different models.

The workflow chart of the study methodology

The CNN and RF models were used to merge data from multiple sources and identify DPISs (Figure 4). First, CNN models were used to train sequential images from Gaofen-2 to specify the differences in features between informal settlements and other settlements. Second, the RF model was used to merge the vectorized nighttime light data, the normalized difference vegetation index (NDVI), RTUD, and vectorized high-resolution image features obtained from the CNN model. Scenario-scale informal settlement identification results were simulated by combining multivariate geographical data. Finally, the DPIS results obtained via RF simulation were overlaid with the urban settlements obtained from the road network to obtain the final identification results. Spatial resolutions of nighttime light, NDVI, and RTUD were unified to 30 m by spatial interpolation during data processing.

Figure 4. Informal settlement identification based on multi-source geographic information data.

Recognition and training of remote sensing images

Convolutional Neural Networks is the first learning algorithm to successfully train multilayer networks based on traditional artificial neural networks. Its main advantage compared with its predecessors is that it automatically identifies relevant features without any human supervision (Alzubaidi et al. 2021). A CNN model is composed of several key layers including an input layer that receives data, a convolutional layer that processes the input data, a pooling layer that down samples the output of the convolutional layer, a fully connected layer that connects all the neurons in the previous layer, and an output layer that produces the final result.

The CNN model was based on the VGG-16 model architecture (Ren et al. 2017) It was primarily used to recognize Gaofen images containing 13 convolutional, 5 pooling, and 3 fully connected layers, taking the input image size of 256 × 256 pixels as an example (Figure 5). The training sets comprising 600 samples were selected from the study area (300 samples for informal settlements and 300 samples for other settlements). The number of hidden units in the FC7 layer was modified from 600 to 60 to balance the contributions of high-resolution and 30-m resolution features since the features extracted from CNN (the hidden units in the second fully connected layer (FC7)) are fused with the 60 original features and utilized in the residential scene classification based on feature fusion and decision fusion. Additionally, the number of hidden units in the last fully connected layer (FC8) was modified from 1000 to 3 since three types of high-resolution sample images were used in our proposed CNN.

Figure 5. Architecture of the VGG16 network.

The datasets were classified at the patient image level into 80% training and 20% testing for the different stages of learning and were randomly separated into 150 batches for every epoch; the number of iterations (epochs) was set at a maximum of 20 from the behavior of validation loss. Fourfold cross-validation was used to evaluate the performance of the current method. This cross-validation method prevents overfitting and guarantees generalization. This process was repeated for each architecture (basic CNN, VGG16 transfer, and VGG16 fine tuning). Building, training, and prediction of the deep-learning models were performed using the Keras library (https://keras.io) and TensorFlow backend engine.

Informal settlements were largely occupied by densely populated shantytowns and urban villages, whose main characteristics included dilapidated roofs of houses (Figure 6a), tightly arranged houses, and small spaces in Gaofen images (Figure 6b). Moreover, the orientation of houses in some informal settlements was confusing owing to their unstructured distribution (Figure 6c). The ratio of the training set to the test set in the network was 5:1, and the obtained test accuracy was 81.45%.

Figure 6. Characteristics of informal settlements in Gaofen images. (a) dilapidated houses. (b) high-density distribution. (c) Disorderly distribution.

Auxiliary geospatial data fusion

Random forest is a type of machine learning that utilizes decision trees to make predictions and is known for its high predicting outcome accuracy, its ability to handle outliers and noise in the data, and its tendency to avoid overfitting. This model was based on the bagging method framework (Breiman 1996) that constructs multiple samples by randomly extracting data from the originals, then uses the random splitting technique of nodes to construct multiple decision trees for each resampled sample. The classification results are determined by the median of the output categories of individual trees, and the classification error is based on both the ability of each tree to classify data and the correlation between them. In our study, the second fully connected layer simulated from the CNN model was used as the input sample with one-dimensional vectorized nighttime lights, NDVI, and RTUD, and was recorded as $\left\{h_1\left(X\right),h_2\left(X\right),h_3\left(X\right)\right\}$. The formula is expressed as follows:

$H\left(x\right)=arg\underset{Y}{max}\sum _{i=1}^{3}I(h_i\left(x\right)=Y)$

K-fold cross-validation revealed that the out-of-bag error of the model was the lowest and tended to be stable when the classified trees were 1000. The ratio of the training set to the test set in this model was 3:1, and the final test accuracy was 91.37%.

Results

Distribution of DPISs and population ratio

The DPISs had distinct patterns of distribution in Beijing, Shanghai, and Guangzhou (Figures 7 and 8). The DPIS in Beijing was mainly located in the southern part of the city. They were contiguous and accounted for 1.98% of the total area. The DPIS in Shanghai was concentrated in urban centers and accounted for 0.67% of the total area. The DPIS in Guangzhou had a more dispersed distribution, occupied a larger area compared with the DPISs of Beijing and Shanghai, and accounted for 3.95% of the total area.

Figure 7. Spatial distribution of densely populated-informal settlements (DPISs) in the study areas: (a) Beijing, (b) Shanghai, and (c) Guangzhou. The background of the map is from Gaode map (https://www.amap.com/).

Figure 8. Densely populated-informal settlements and formal settlement (FS) images obtained from Gaofen-2 satellite images.

The population numbers and densities of DPISs were obtained in the main urban areas of Beijing, Shanghai, and Guangzhou by overlaying the WorldPop population data with the identified results (Figure 9). Shanghai had the smallest DPIS and the highest population density at 42,800 people/km². Beijing had a larger area of DPISs than Shanghai, although its total population and population density was the smallest at 7,100 people/km². The DPIS of Guangzhou is the largest and has the highest total population, (20,800 people/km²).

Figure 9. Areas, total population, and density statistics of DPISs in the study area.

Analyzing the relationship between urban land-use change and DPISs

China entered a phase of rapid urbanization starting in 1980. The urbanization rate of Beijing and Shanghai exceeded 85%, and that of Guangzhou exceeded 70% by 2018 (Yu et al. 2019). The calculations in our study were derived from two years of land-use remote sensing monitoring data released by the Resource Environment Science and Data Center, Institute of Geographical Sciences and Resources, Chinese Academy of Sciences: 1980 and 2020. There was a significant amount of rural land that was transformed into urban areas owing to urbanization, with urban areas in Beijing expanding from the center to the periphery, urban areas in Shanghai mainly expanding to the northwest, and urban areas in Guangzhou mostly expanding to the north and east (Figure 10).

Figure 10. Urban land-use changes and DPIS distribution in the study areas from 1980 − 2020: (a) Beijing, (b) Shanghai, and (c) Guangzhou.

The land-use changes in DPIS areas of the three cities were determined (Figure 11). Over 60% of the DPISs in Beijing were converted from rural residential land to urban land in 1980, primarily existing in the form of urban villages, and most were located in the southern part of Beijing city. Over 80% of Shanghai’s DPISs were found in old urban areas, and less than 8% were found in the urban villages around the city. The urbanization rate of Guangzhou was not as high as that of Beijing and Shanghai, although over 35% of the DPIS were distributed in rural and other built-up areas, and over 30% of the area was converted from rural to urban land.

Figure 11. Comparison of land-use changes in DPISs in three cities: Beijing, Shanghai, and Guangzhou.

Quality validation of DPIS regional settlement

The feasibility of the experiment was confirmed by selecting 1,500 and 2,500 sampling points in FSs and DPISs, respectively, and comparing the distribution of NDVI, nighttime lights and RTUD in Beijing, Shanghai and Guangzhou (Figure 12). The FSs comprised of villa mixed residential units, high-rise residential (HRR) units, affordable housing (AH), and garden villas (GV), whereas the DPISs were low-income residential areas, old public houses, and historical areas (HA). Overall, the DPISs had a lower NDVI, stronger nighttime lights, and higher RTUD compared with FSs. This indicated that most DPISs were of lower quality than that of FSs, and were highly densely populated with little green coverage. The characteristics of different settlements were further analyzed.

Figure 12. Distribution of normalized difference vegetation index (NDVI), night lighting, and real-time Tencent user density (RTUD) in different types of settlements in (a) Beijing, (b) Shanghai, and (c) Guangzhou.

The DPISs showed a lower green coverage compared with FSs, and the distribution interval was narrower. Among them, 80% showed an NDVI of less than 0.4. However, the HA in Shanghai was of a special type and showed a higher NDVI than that of residential areas such as AH and GV. The HA in Shanghai was originally built in the 1920s and 1930s as an office of Public Concession Patrol House, a shelter for Public Security Bureau staff; it was classified as an immovable cultural relic in 2004 owing to its historical legacy. We visited some DPISs in the study area, including Wanping City in Beijing, Shikumen in Shanghai and the riverside area in the south of Huangpu District in Guangzhou. The lack of parks and street trees around DPISs was observed in the field study (Figure 13).

Figure 13. Field investigation photo records and Baidu Street Views of DPISs in 2019. (a) and (b) Wanping city, Beijing. (c) and (d) Shikumen, Shanghai. (e) and (f) South side of Huangpu District, Guangzhou.

Nighttime lights and RTUD can indicate the traffic flow after 21:00 and 2:00 respectively, reflecting the population density at night. Nighttime lights in DPISs were stronger than those in FSs in Shanghai and Guangzhou. The opposite was found in Beijing since DPISs were mainly located in the peri-urban areas where the construction of urban infrastructure lags behind, resulting in lower nighttime light density. In contrast, Shanghai and Guangzhou have experienced long-term urbanization and economic growth since they developed earlier in terms of urban development. The DPISs in these cities may be located in the city center owing to historical reasons or urban expansion. These areas typically have higher economic activities and population density. The RTUD of DPISs were higher than that of FSs among the three cities, especially in Shanghai and Guangzhou.

The NDVI, nighttime lights, and RTUD could reflect the characteristics of DPISs, but a single data description was one-sided, and these elements needed to be collectively considered. For example, HRR units have a large number of floors leading to a large population, but their infrastructure was more developed, and the vegetation coverage did not fit the characteristics of DPISs. Field investigations confirmed this conclusion. Photographic records and questionnaires were used to illustrate the housing, demographic, economic and transportation characteristics of the study areas. The per capita living area of DPISs ranged from 0‒30 m². The residents’ satisfaction level was low, and the main problems were dilapidated houses, few green areas, noise, and lack of security issues, among others.

Discussion

Similarities and differences between DPISs and SDG 11.1

The types of settlements mentioned in SDG 11.1 are slums, informal settlements, and inadequate living spaces. Slums and informal settlements share common characteristics, such as limited access to clean drinking water, inadequate sanitation facilities, poorly constructed housing units, and a lack of legal ownership or tenure according to UN-Habitat (Krizhevsky, Sutskever, and Hinton 2017; Shorten and Khoshgoftaar 2019; Zhang et al. 2017). However, informal settlements exclude crowded living spaces while slums include them since people living in informal living spaces usually have multiple income levels – not all of them are residents with income levels below the poverty line. Inadequate living space is primarily measured in terms of equitable living conditions (Eide 2018) and is characterized by a lack of legal residency rights, affordability, demand for special populations, livability, adequate infrastructure, and ample culture; this is quite different from slums and informal settlements (Kohli et al. 2012). The DPISs identified in this study reflected the rapid urbanization process in China together with planning of the cities. They were located in densely populated older settlements whose housing characteristics were mainly informal in the form of dilapidated roofs, tightly packed houses, and small common spaces. Combined with the field survey, people living in DPISs were mainly foreign tenants living in poor housing environments and insufficient dwelling space. A comparison of slums, informal settlements, and inadequate living spaces defined in SDG 11.1 (SDG Indicator Metadata 2021) showed considerable similarities and differences (Table 2).

Table 2. Characteristics of settlements defined in sustainable development Goal (SDG) 11.1 and densely populated‒informal settlements (DPISs).

Characteristics	Settlements in SDG 11.1			DPIS
	Slums	Informal settlements	Inadequate living space
Lack of quality drinking water	√	√	-	×
Lack of improved sanitation	√	√	-	○
Crowded living space	√	-	-	√
Residential structures of questionable quality	√	√	-	○
Lack of legal residency rights	√	√	√	○
Lack of affordability	-	-	√	○
Lack of demand for special populations	-	-	√	○
Lack of livability	-	-	√	√
Lack of adequate infrastructure	-	-	√	√
Lack of adequate culture	-	-	√	○

Notes: “√” indicates compliance, “×” indicates noncompliance, “○” indicates partial compliance, and “-” indicates that this characteristic is not included.

Crowded living space (densely populated areas), lack of livability (small public spaces), and inadequate infrastructure (such as lack of green spaces, parks, and recreational facilities) had the same characteristics of DPISs and settlements as defined in SDG 11.1 (Wu, Zhang, and Webster 2013). Notably, the lack of quality drinking water had the most remarkable difference. Urban areas in China have largely achieved universal access to quality tap water, and its supply is also available to DPISs. However, a lack of improved sanitation, quality housing structures, legal tenure, affordability, demand for special populations, and adequate culture was only partially observed in our study.

Some DPISs identified in Shanghai (called lane buildings) were primarily located in old urban settlements classified as historic districts and remained as local tourist attractions. They had certain basic requirements for sanitary conditions that were improved. However, urban villages in DPISs (especially in Guangzhou and Beijing) were at risk of disease transmission owing to poor sanitation and lack of management. Similar conditions were found in “Residential structures of questionable quality.” Indigenous people converted their village houses into multi-story buildings without any authority in some DPISs; part of the illegal self-built buildings were of poor quality and did not meet the standard requirements (Yu, Xie, and Chan 2019). However, historically preserved buildings (such as lanes) offered good residential structures.

“Lack of affordability” and “Lack of demand for special populations” also need to be discussed in two cases. Locals built illegal residential structures on the backdrop of original houses or took up vacant fields such as farmlands in the suburban areas to build unauthorized dwellings to accommodate the increasing population with the expansion of cities. Most of the population living here had no legal residency rights, except for some houses that were rented as stipulated. The Chinese government introduced policies to control the illegal expansion of urban houses and to protect the spatial distance between buildings in the 1980s. However, they were not strictly enforced, with many suburban areas retaining illegal houses that were built without government authorization (Wang, Wang, and Wu 2009; Hao, Sliuzas, and Geertman 2011). Not all residents living in DPISs were unable to afford a house elsewhere; some only lived there as temporary transitional housing to save on the cost of living owing to the low rents and convenient transportation to the surrounding metropolitan area.

The special population group primarily refers to the old, weak, diseased, disabled, pregnant, and other groups with inferior health or physical conditions compared with healthy people. These groups tend to require more care and attention in all aspects of housing, traveling, and living. Some DPISs were well-connected and accessible with low-rise housing that was conducive to special populations. However, most had problems, such as chaotic traffic order, insufficient road space, and prominent parking conflicts (Li, Kleinhans, and van Ham 2018). We conjecture that poor housing conditions do not affect the creation of cultures, and many people may rent houses in DPISs owing to their cultural associations.

Role of DPIS in evaluating urban sustainability

Multi-source geographic information data and machine learning were used to identify the distribution of DPISs and population ratios in the urban areas of Beijing, Shanghai, and Guangzhou and determine its role as a localization indicator for SDG 11.1 in China. Considerable differences existed in the distribution of DPISs and population ratios within the main urban areas of Beijing, Shanghai, and Guangzhou. Shanghai’s DPIS was concentrated in the center of the city, with the smallest area and the largest population share. Beijing’s DPIS was located in the southern part of the city, with contiguous areas and the smallest population share. Guangzhou’s DPIS was more scattered and had the largest area. We concluded that Shanghai is likely to be more intense, broader, and more difficult to manage in the event of a sudden disaster than Beijing and Guangzhou owing to its central concentration of the DPIS and high population density. Megacities with high population density were more likely to have major outbreaks with most of them starting in DPIS areas without management and control compared with small and medium-sized cities with sparse populations as observed during the outbreak of coronavirus disease 2019 (COVID-19). This was proved by the outbreak of COVID-19 in Shanghai in March 2022. Sethi and Creutzig (2021) suggested that COVID-19 impacts are especially detrimental in urban centers in large cities that house some of the densest communities on the planet, and pandemic-resilient cities require rental-housing stocks and highly accessible urban environments.

There was a clear relationship between the existing urban layout and urbanization. Considerable differences existed in the composition of DPISs within the main urban areas of Beijing, Shanghai, and Guangzhou. The land-use changes from 1980–2020 showed that Beijing’s DPISs were mainly composed of urban villages, Shanghai’s DPISs were mainly composed of old settlements distributed in the urban center, and Guangzhou’s DPISs were more diverse, occupied by urban villages, old settlements, and irregular rural settlements. Most DPISs developed from urban villages among the first-tier cities in China, except for the historical buildings in Shanghai. They had poor sanitation and lacked infrastructure, but tended to have good greenery since these urban villages were located in urban-rural areas. Therefore, DPISs cannot summarize the current status of a city’s development.

SDG 11.1 is a comprehensive indicator that should be considered in the context of multiple dimensions when applied to specific countries and regions. A single quantification of area cannot express the degree of sustainability. The settlements identified in our study shared some of the same characteristics compared with SDG 11.1, in which slums, informal settlements, and inadequate living space were measured: congested living space, lack of livability, and inadequate infrastructure. These characteristics are an important indicator of the sustainable development of Chinese cities. However, they need to be considered in conjunction with other elements when making an evaluation.

Therefore, DPISs and population ratios can be used as localized indicators for SDG 11.1 in China to effectively evaluate the sustainability of cities. However, decision-makers should consider the differences between cities based on their development history and locational environment to develop effective sustainable development policies (Jiang et al. 2021). Maintaining a balance between DPISs and population ratios, improving the urban DPIS environment, and reducing the DPIS area by an appropriate amount can ensure healthy urban development and stronger resistance in case of major disasters. The existence of DPISs involves various elements, such as economic, social, environmental, transport, and cultural aspects. Improving the well-being in urban informal settlements can effectively strengthen urban shortcomings. At the same time, there is an opportunity for public health, public administration, international aid, non-governmental organizations, and community groups to innovate beyond disaster response and move toward long-term sustainable and resilient planning (Corburn et al. 2020), especially under deep uncertainties related to climate change (Haasnoot et al. 2013; He et al. 2021; Lempert, Popper, and Bankes 2003).

Conclusions

Based on multi-source geographic information data and machine learning, the present study identified the distribution of DPISs and population ratios in the three urban areas of Beijing, Shanghai, and Guangzhou, and discussed their role as a localization indicator for SDG 11.1 in China. The main conclusions were as follows:

1. Significant differences existed in the distribution of DPISs and population ratios within the main urban areas of Beijing, Shanghai, and Guangzhou. Shanghai’s DPISs were concentrated in the center of the city, with the smallest area and the largest population share. Beijing’s DPISs were mainly located in the southern part of the city, with contiguous areas and the smallest population share. Guangzhou’s DPISs were more scattered and had the largest area.

2. Significant differences existed in the composition of DPISs within the main urban areas of Beijing, Shanghai, and Guangzhou. Land-use changes from 1980 to 2020 showed that Beijing’s DPISs were mainly composed of urban villages, Shanghai’s DPISs were mainly composed of old settlements distributed in the urban center, and Guangzhou’s DPISs were more diverse and occupied by urban villages, old settlements, and irregular rural settlements.

3. The distribution of settlements and population ratios identified in this study can be used as localization indicators for SDG 11.1 in China. The settlements identified in this study shared some same characteristics compared with SDG 11.1, in which slums, informal settlements, and inadequate living space were measured: congested living space, lack of livability, and inadequate infrastructure. This study had more safety hazards that reflects the rapid urbanization process in China and the city’s inner planning. This serves as an important indicator of the sustainable development of Chinese cities.

In conclusion, this study provides valuable insights into the identification and analysis of DPISs within urban areas in China. However, there are several potential directions for future research that can further enhance our understanding of DPISs and their implications for urban development and sustainability.

First, conducting fine-grained analyses can provide a deeper understanding of DPISs by exploring specific characteristics and subtypes, such as variations in housing conditions, infrastructure provision, and socioeconomic factors within these settlements. Second, longitudinal analyses using multiple time points will offer insights into the dynamic nature of DPISs. Researchers can identify trends, patterns, and shifts in the spatial distribution of DPISs by tracking their growth and changes over time. Additionally, conducting comparative analyses across different cities or regions within China and internationally will enable researchers to uncover similarities and differences in the characteristics, drivers, and impacts of DPISs. Finally, researchers can provide evidence-based recommendations to policymakers and urban planners by evaluating the effectiveness and impact of policies, interventions, and urban planning strategies targeted at DPISs.

Addressing these future research directions will deepen our understanding of DPISs, their dynamics, and their implications for sustainable urban development. Ultimately, this knowledge will facilitate the design of interventions and policies that enhance the well-being and quality of life for individuals residing in DPISs.

Data and codes availability statement

Datasets used:

1. Gaofen-2: open dataset at http://www.cresda.com/CN/index.shtml

2. Luojia-1: open dataset at http://59.175.109.173:8888/

3. Sentinel-2: open dataset at https://scihub.copernicus.eu/

4. RTUD: open dataset at https://heat.qq.com/index.php

5. Road network data: open dataset at http://www.openstreetmap.org/

6. WorldPop dataset: open dataset at https://www.worldpop.org/

7. China Land Use Remote Sensing Monitoring Data: open dataset at https://www.resdc.cn/-data.aspx?DATAID=335

8. DPIS dataset: https://figshare.com/articles/dataset/Untitled_Item/23805075

Acknowledgments

We thank many students (including Jingbo Wang, Yi Du, Xun Yu, and Kunqin Liu) from the Department of Geography and Environmental Sciences, Shanghai Normal University for their hard work in processing remote sensing images and extracting densely populated‒informal settlements.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the National Natural Science Foundation of China under Grant numbers 42171389 and 41730642.