Revealing the Impact of Urban Form on COVID-19 Based on Machine Learning: Taking Macau as an Example

Abstract

The COVID-19 pandemic has led to a re-examination of the urban space, and the field of planning and architecture is no exception. In this study, a conditional generative adversarial network (CGAN) is used to construct a method for deriving the distribution of urban texture through the distribution hotspots of the COVID-19 epidemic. At the same time, the relationship between urban form and the COVID-19 epidemic is established, so that the machine can automatically deduce and calculate the appearance of urban forms that are prone to epidemics and may have high risks, which has application value and potential in the field of planning and design. In this study, taking Macau as an example, this method was used to conduct model training, image generation, and comparison of the derivation results of different assumed epidemic distribution degrees. The implications of this study for urban planning are as follows: (1) there is a correlation between different urban forms and the distribution of epidemics, and CGAN can be used to predict urban forms with high epidemic risk; (2) large-scale buildings and high-density buildings can promote the distribution of the COVID-19 epidemic; (3) green public open spaces and squares have an inhibitory effect on the distribution of the COVID-19 epidemic; and (4) reducing the volume and density of buildings and increasing the area of green public open spaces and squares can help reduce the distribution of the COVID-19 epidemic.

1. Introduction

1.1. Research Background

As of 16 August 2022, since the outbreak of global transmission in early 2020, the total number of reported cases of COVID-19 was 590 million, affecting 7% of the world’s population and 210 countries, with a mortality rate of 1.09% [1]. According to global anti-epidemic research over the past two years, various studies have found that the spread of COVID-19 is related to many factors: housing quality and living conditions, crowding, regional climate, air pollutants, population migration, and government intervention [2,3,4,5]. At present, the methods for studying COVID-19 from an epidemiological perspective are mainly traditional kinetic models and statistical models: classic compartmental models, SIR [6], SIER [7], SEIRS [8], and SEIHR [9]. They rely on people flow data for core analysis. Therefore, in the use of traditional models, less consideration is given to the information elements of material space. With the development of artificial intelligence technology, machine learning methods can be used to analyze the characteristics of the COVID-19 epidemic, so as to achieve efficient epidemic prevention and control [10].

1.2. Literature Review

Machine learning techniques have been widely used in infectious disease research, including severe acute respiratory syndrome (SARS), H1N1 influenza virus, and Middle East respiratory syndrome coronavirus (MERS-CoV) [11]. Currently, machine learning-based COVID-19 research can be divided into the following four categories: (1) epidemiological causal inference of nonlinear and intervariable interactions and complex processing of multidimensional data [12,13]; (2) disease prediction, diagnosis, prognosis, and clinical decision making [14,15,16,17,18]; (3) the use of random forests to identify and analyze diseases, so as to conduct genome-wide association studies [19,20,21]; and (4) spatial epidemiological research based on the combination of geospatial information and remote sensing data, including the meteorological data and case distribution map [22], the urbanization and COVID-19 vulnerability distribution map [23], and the distribution map of predicted transmission [24,25]. It is clear from the literature that the first three categories are currently the most widely used, especially in the medical field. The research direction of spatial epidemiology needs to be further explored, which also depends on the further mining of geographic information data in the future.

1.3. Problem Statement and Objectives

Most of the machine learning-based research on the COVID-19 epidemic focuses on the prediction of the COVID-19 virus, with fewer studies on the impact of urban environmental factors on the epidemic. In this paper, the sample of the study is improved. First, the COVID-19 virus hotspot distribution map is used as training set A, and the city morphology map is used as training set B. The ultimate goal is to use the distribution of the COVID-19 virus in cities to predict urban form, so as to deeply study the impact of urban form on the COVID-19 virus. Second, a conditional generative adversarial network (CGAN) is employed to analyze the relationship between urban spatial risk factors and urban form. Lastly, assuming different risk distribution patterns, the effects of different urban patterns on the COVID-19 virus are analyzed.

This study proposes an image-based CGAN, so that the COVID-19 virus heat map can be used as a material for predicting urban morphology maps, and the research results can reflect the promotion or inhibition effect of urban morphology on the COVID-19 virus. This research process can also be used in other studies related to epidemics and urban forms. The research process is as follows: (1) taking Macau as a case study, statistics of the transmission trajectory of new coronavirus patients in the city, as well as a COVID-19 hotspot map are established as the material for machine learning; (2) in the area corresponding to the COVID-19 hotspot map, different colors are used to extract and simplify the content of the urban form, and it is used as secondary material for machine learning; (3) the above materials are used as input into the CGAN and the generated weight model is tested. Furthermore, the distribution of the COVID-19 epidemic in other regions of Macau is imported for forecasting; (4) assuming the distribution of the COVID-19 epidemic to different degrees, including extreme cases of complete distribution and no distribution, the derivation of urban form is carried out; and, lastly, (5) the above test results are compared and analyzed, and the impact of urban form on the distribution of the COVID-19 epidemic is summarized.

2. Materials and Methods

2.1. Study Area and Data Sources

This research is based on the image synthesis technology of machine learning, and an urban morphology map is generated through the footprint heat map of the COVID-19 epidemic. The physical patterns, layouts, and structures that make up an urban center are collectively called the urban form. The study of urban morphology is an indispensable part of studying the form of human settlements, the process of their formation and transformation, and urban planning and design, and it helps to understand and analyze the process and characteristics of urban development. According to a book written by Vitor Oliveira, urban morphology is the science that studies the physical form of cities, as well as the main agents and processes shaping them over time [26]. In this study, the main urban form elements considered are roads, green spaces, water/coastline, buildings, and vacant land.

First, the heatmap of the COVID-19 epidemic footprint is used as training set A, and the corresponding urban morphology map of Macau Peninsula (Macau is an inalienable part of China’s territory, consisting of the Macau Peninsula, Taipa, and Coloane, with a land area of 32.9 square kilometers. The Macau Peninsula is connected to mainland China (Zhuhai City, Guangdong Province) to the north, and to Taipa to the south by the Ponte Governador Nobre de Carvalho (Carvalho Bridge), the Ponte de Amizade (Friendship Bridge), and the Ponte de Sai Van (Sai Van Bridge). Taipa and Coloane are connected by a 2.2 km-long, six-lane highway.) is used as training set B. Then, a conditional generative adversarial network (CGAN) is implemented for training [27]. In the image translation using training set A and training set B, the generator and the discriminator are allowed to play against each other, thus improving the quality of the generated pictures and realizing the ability to generate urban morphology maps [28].

The experimental materials are shown in Figure 1. In the processing of the Macau map, in order to simplify the data, various elements in the map were represented in different colors and presented in the form of color pictures [29]. In this study, five colors were used to represent the elements on the map: roads and squares are red (R = 255, G = 0, B0), green spaces are green (R = 200, G = 215, B = 158), water is blue (R = 158, G = 188, B = 216), buildings are white (R = 255, G = 255, B = 255), and land is black (R = 0, G = 0, B = 0). These five colors represent most of the content in the city map. The footprint hotspot data of the COVID-19 epidemic were generated by researchers from the statistics of the footprint report of a total of 500 patients in Macau (from mid-June to early July 2022), which was fully disclosed by the Macau Health Bureau. The addresses of the footprints were mainly registered according to the building of residence. Despite the longest residence time and the highest risk of carrying the virus, due to personal privacy, some private itineraries were not officially announced. Therefore, this study could only screen the footprints of a total of 3265 confirmed patients on the Macau Peninsula (more details can be found in Appendix B, Table A1). Then, the addresses were converted into latitude and longitude coordinates using Google Maps API Web Services, before being input into ArcGIS Pro to generate hotspot data. Since CGAN requires paired datasets for training, in order to make the data correspond one-to-one, they were uniformly corrected into the Observatorio Meteorologico 1965 Macau Grid.

Figure 1. Experimental research materials and study area.

2.2. Model Construction Process

Since machine learning requires a large number of samples, in order to obtain more accurate experimental data, the sample images were divided into grids with a size of 512 × 512 pixels, and each image slice was about 4 (ha) in area. After weighing the quality and quantity, a 6 × 6 grid was obtained, and 36 images of the urban morphology map, the COVID-19 epidemic heat map, and the hypothetical heatmap were cut into 36 images, yielding a total of 144 samples (Figure 2).

Figure 2. Research samples.

The conditional generative adversarial network (CGAN) is a variant of the generative adversarial network (GAN). Consistent with the original GAN, the CGAN is mainly composed of two adversarial models: a generator responsible for generating images and a discriminator for judging the authenticity of the generated images. As shown in Figure 3, the main principles are as follows: (1) the generator generates fake pictures according to the input picture (Train A) and random vector (Z); (2) the discriminator determines another set of corresponding pictures (Train B) and random vectors as true pictures. At the same time, it is compared with the fake pictures input by the generator, whereby the real pictures are marked as 1, and the fake pictures are marked as 0; (3) if the generated image is judged to be false, the discriminator returns the deviation value between the fake image and the real image to the generator. The generator is subsequently upgraded so that it can generate more realistic pictures. On the contrary, if the discriminator judges that the generated image is real, the discriminator continues to learn from the training set to improve the recognition ability; and (4) through adversarial training, the generator can finally generate fake and real pictures, so as to achieve the goal of generating urban morphology maps.

Figure 3. CGAN principle.

3. Results, Analysis, and Discussion

3.1. Training Result

Upon training the model for multiple iterations, we found that the loss values of the trained generator and discriminator fluctuated significantly but tended to decrease overall. In Figure 4, the orange line represents the fluctuation curve in the machine learning process, i.e., the stability of the learning result. It should be noted that the stability of machine learning does not represent the accuracy of the learning results of the model. Different input conditions and the completeness of input data have different effects on the stability of machine learning results, and the correlation between learning conditions and target results can also be indirectly reflected.

Figure 4. Loss values for during training.

The different input data in Figure 5 were taken as an example, while the COVID-19 epidemic distribution hotspot was taken as the input data. The stability of the model results after 200 iterations of learning revealed a certain correlation between the distribution of the COVID-19 epidemic footprint and the components of urban form.

Figure 5. Model training process.

At the same time, Figure 5 shows the comparison of iterative training results for model learning accuracy. The hotspots of the epidemic distribution, the actual urban form, and the derived urban form results (horizontal columns) were iteratively learned 50 times, 100 times, 150 times, and 200 times. The “input” in each group represents the spatial data of the distribution of COVID-19 epidemic hotspots in the target area. “Real” represents the urban form distribution of the target area, which is only used as a reference for comparison of results and does not participate in machine learning training. “Generated” means that the machine has generated a prediction result on the urban form distribution of the target area.

It can be seen that, under the condition of 50 iterations, the learning results of the machine were blurred, and the accuracy was lower than that of the real urban form area. Under the condition of 150 iterations, the results of machine learning improved, but the accuracy was still not ideal. Under the condition of 200 iterations, it can be seen that the similarity between the urban form area obtained from the basic learning results and the actual urban form area reached a high level. At the same time, is also shown that, under the condition of maximizing the saving of machine load and learning time cost, improving the accuracy of this machine learning model could basically meet the requirements of the target after 200 iterations.

With the increase in training, the study also found that: (1) the overall accuracy of training significantly improved; (2) in the urban area, the distribution of roads did not improve significantly; and (3) the range of buildings could be deduced from the distribution of COVID-19 epidemic hotspots. Compared with real urban areas, a high degree of similarity was restored, but the accuracy of building blocks and shapes was not significantly improved. For example, in the 200th iteration, the real urban form had a more regular and cornered road segmentation. However, in the urban form regions derived from machine learning training, the roads were more vivid and curved, although the extent of the buildings was roughly the same.

3.2. Results Comparison of Different Types of Urban Forms

Furthermore, this study took the three urban form areas as three typical categories for comparative analysis. As shown in Figure 6, A2, B2, and C2 are the urban areas formed in three different periods in the city. A2 is an area with many residential and industrial buildings distributed in the 1990s, including large-scale buildings. B2 is an area with several mixed commercial and residential buildings from the 1970s. At present, there are no large buildings, instead showing the characteristics of a mixture of medium and small buildings, representing a typical commercial center that has entered modern society. C2 is the area under the scope of World Cultural Heritage protection, representing the most prosperous streets and commercial centers in Macau in the 1920s and 1930s, including low-rise buildings with smaller volumes. At the same time, it also retains the most traditional and primitive urban form of the city.

Figure 6. Results comparison of different types of urban forms. A1, B1, and C1 are the actual epidemic distribution maps; A2, B2, and C2 are the different types of urban morphology maps in Macau; A3, B3, and C3 are the results predicted by the model of A1, B1, and C1.

Through the training results, it can be found that: (1) the distribution of important urban roads (the widest roads) in the target area had little correlation with the distribution of COVID-19 epidemic hotspots (Figure 6A1–A3); (2) the arrangement of building volumes had a certain correlation with the distribution of COVID-19 epidemic hotspots. Buildings with large volumes highly overlapped with the distribution of COVID-19 epidemic hotspots (Figure 6B1–B3); and (3) when the base area of buildings presented similar plots, the number of buildings in the plot did not affect the distribution of COVID-19 epidemic hotspots (Figure 6C1–C3).

3.3. Model Application and Analysis

The city of Macau consists of three originally independent islands, namely the Macau Peninsula, Taipa, and Coloane, through land reclamation. From the perspective of urban development and intensity, compared with the relative lag in Coloane’s construction, the island and Taipa districts have the same high-density, large-scale, and strongly enclosed street market appearance. The current connection model for the island was able to clearly reflect the internal connection between the main distribution of the epidemic and the urban form. Therefore, the known epidemic distribution points in Taipa were selected as new information to be implanted into the model, which could reflect the prediction and judgment of the urban form of Taipa made by the connection model.

From the analysis of the results generated by the model operation, it was found that the density of the epidemic distribution points was closely related to the objective factors of the city: (1) according to the epidemic distribution shown in A1 in Figure 7, when it is concentrated in a single location, it is often proportional to the construction intensity of that location. As shown in A2 in Figure 7, there are many dense buildings, and the arrangement of undeveloped vacant land along the road and the built environment forms a spatial transition and partition, effectively controlling the epidemic in a specific area. (2) Furthermore, when the epidemic situation is distributed in multiple places and multiple points, as shown in B1 of Figure 7, it shows the urban form generated by B2 of Figure 7. The main reason is that the convenience of road traffic organization is improved, the accessibility between built environments is enhanced, and the degree of enclosure is high, resulting in the rapid mobility and large coverage of the epidemic, which has the greatest impact on the city. (3) Additionally, in areas where the urban road network is concentrated, road intersection squares and surrounding open spaces (black plots) can effectively slow or prevent the spread of the epidemic when the epidemic presents a single sporadic distribution, as shown in Figure 7C1. At this time, the size of the building becomes an important indicator of the degree of impact of the epidemic. (4) Moreover, when the epidemic situation is distributed locally, as shown in D1 of Figure 7, although the built environment is complex, dense, and large in number, as shown in D2, the weakening of urban road accessibility becomes a key factor in the inability of the epidemic to have a large-scale and high-concentration impact.

Figure 7. The result of deriving the urban morphology from the heat map of the epidemic distribution in Taipa. A1, B1, C1, and D1 are different slices of the actual epidemic distribution in Taipa, Macau. A2, B2, C2, and D2 are the results predicted by A1, B1, C1, and D1 through the model.

As analyzed above, the correlation model between urban morphological elements and the impact of the epidemic inferred from the epidemic distribution in Taipa showed that urban road accessibility, built-up environment density, and appropriate land space have an important impact on the spread of the epidemic. Therefore, adjusting the relative relationship among urban morphological elements has a positive effect on epidemic prevention and control.

3.4. Assuming the Epidemic Distribution to Derive the Results of Urban Form

Lastly, images of the distribution of COVID-19 outbreaks with different shapes were assumed in this study. Then, machine learning was used to deduce and generate the urban form. In Figure 8, gray is the hypothetical COVID-19 epidemic distribution area (A1, C1, D1, E1, and F1), white is the presumed distribution area (B1) when the COVID-19 epidemic peaks, and black is the area where no COVID-19 outbreak is assumed. Meanwhile, A1 to F1 represent different distributions that may exist when the COVID-19 outbreak occurs. A2 to F2 are urban-form areas derived from machine learning.

Figure 8. Assumption of the epidemic distribution to derive the results of urban form. A1, B1, C1, D1, E1, and F1 are the researchers’ assumptions about the distribution of different outbreaks. A1 and B1 represent different epidemic distribution intensities. C1, D1 represent different distribution shapes of the epidemic. E1 and F1 represent different positive and negative shapes of the epidemic distribution. A2, B2, C2, D2, E2, F2 are the results of A1, B1, C1, D1, E1, and F1 predicted by the model.

The results of the study found that: (1) urban form is related to the distribution of the epidemic, but it has a weak relationship with the intensity of the distribution of the epidemic; (2) furthermore, urban form is related to the shape of the epidemic distribution, and the results of the derived urban form are roughly consistent with the scope of the epidemic distribution area in the outline; and (3) areas with more epidemic distribution have a higher building density. In areas with fewer outbreaks, buildings are more sparsely distributed.

4. Discussion: Pandemic and Sustainable Living

In order to combat the wealth gap, climate change, gender equality, and other issues, in 2015, the United Nations launched the “2030 Sustainable Development Goals” (SDGs), proposing 17 core goals for global governments and enterprises to jointly move towards sustainable development. SDG Goal 11 is “building cities and villages that are inclusive, safe, resilient, and sustainable”. On 9 July 2020, UN-Habitat and the World Health Organization jointly hosted an online forum on “Urban Form and COVID-19: Reflections on Density, Overcrowding, Public Space, and Health” [30,31]. In the current context of the spread of COVID-19, how can local governments take action to implement the United Nations 2030 Agenda for Sustainable Development? Moreover, how should government officials, professionals, and scholars better understand the relationship between urban form and disease transmission and prevention? Perhaps, this is one of the issues we need to think about in regard to how to sustainably develop urban life. The reality of facing COVID-19 is that, in low-income neighborhoods, developers have no incentive to increase floor space or require additional infrastructure improvements. Especially in some high-density cities, people live in tighter quarters, often in multigenerational households, and they work in jobs that require face-to-face interaction. The risk of contagion increases as communities lack physical structures and amenities to enhance livability, and residents have no choice but to go out every day to find work or services.

Combining machine learning with the results of the relationship between the COVID-19 heat map and urban form to shape sustainable urban life, the following suggestions can be made in the field of planning and design:

(1) Reconsider the scale, design, and spatial distribution of public spaces. Public spaces can help reduce the risk of spreading COVID-19. Due to the long-term stay-at-home orders, people began to seek physical recovery and psychological pressure relief from green spaces, and the demand for urban green spaces has also increased. However, in the face of cities of different scales, accurate data are required for the allocation of public spaces in order to achieve a better balance of community resources.

(2) Attach importance to the design of green open space. In this study, it was found that there were few or no outbreak clusters in the green space distribution area. It is also possible that the ecological purification effect brought by the plant landscape on green land can effectively slow the spread of the epidemic. At the same time, people enter the green open space to get physical exercise, further improve their physical and mental health, and incorporate an auxiliary role in resisting diseases. Therefore, in urban planning, it is also necessary to consider the combination of various types of green open space (pocket parks, terrace recreation areas, atria of high-density buildings, and rooftops of commercial buildings).

(3) Avoid large-scale architectural designs in the development of residential areas. A single, large building can easily cause crowds. Once a danger occurs, it is easy to cause safety hazards if not evacuated promptly and effectively. Therefore, from architectural planning, a more scattered, multibuilding, and organically combined design mode can be considered. At the same time, attention should be paid to fully reserving space for disaster prevention and emergency response in architectural design. At the same time, the design and construction process should be full of “elasticity” and “resilience”, so as to comprehensively improve disaster prevention and mitigation capabilities.

(4) Pay attention to the semi-enclosed building combination. COVID-19 spreads via aerosols. However, many shopping malls today are introverted “shopping boxes”. In the face of major epidemics, this has instead become a weakness. The primary consideration for consumers is how to strengthen their own protection in this space. Therefore, some lifestyles are gradually changing, and some businesses with open spaces and better ecological environments are deliberately selected for consumption. Open blocks not only disperse the flow of people, but also prevent the accumulation of dense spaces. The advantages of openness, better ventilation, and more outdoor space make it easier for these areas to become the first choice for leisure consumption. Promoting the development of greenway commerce, park block commerce, and “park+” urban commercial complexes, and integrating the concept of green and sustainable living with architecture and commerce are also design models for the current response to the COVID-19 epidemic.

5. Conclusions

Through machine learning, this study used the heat map of the distribution of the COVID-19 epidemic in Macau to derive the urban form, and the following conclusions can be drawn:

(1) Through CGAN, the distribution area of the COVID-19 epidemic can be used to deduce urban forms that may be high-risk and prone to epidemics. This method has potential applications and practical value in the field of future urban design. The relationship between the spatial form of urban risk-prone areas and the degree of epidemic distribution is predicted by the model, and urban design iteration is carried out, which also has certain universality and reference in other cities.

(2) From the results of model training and model application, it can be seen that, when the urban epidemic distribution heatmap was used as the input data for learning, the stability of the model learning results was poor, but the accuracy gradually improved. Under the consideration of saving machine load and learning time cost, the prediction accuracy of the model after 200 iterations of learning and training could basically meet the requirements of target prediction.

(3) From the comparison of the epidemic distribution heat map, the actual urban form area, and the derived urban form area, it can be seen that the combination of urban forms is related to the risk of epidemic occurrence. Larger buildings have a high degree of overlap with the distribution of COVID-19 epidemic hotspots. Areas with a high degree of road enclosure highly overlap with the distribution of COVID-19 epidemic hotspots. These two types of urban forms require special attention. Green public open spaces and squares have an inhibitory effect on the distribution of the COVID-19 epidemic. Reducing the building volume and density can not only increase the area of green public open space and squares, but also help reduce the distribution of the COVID-19 epidemic.

The spread of the COVID-19 outbreak has caused the public to rethink the issue of public health governance. At the same time, it also allows government departments, planners and architects, and experts and scholars to rethink the sustainable development of urban decision making. Needless to say, in addition to epidemic spread, policymakers and planners must consider many other factors when considering residential density, such as economic thresholds and dynamism, social mix and dynamism, urban sprawl, and per capita infrastructure costs. The impact of building density and urban form is part of the comprehensive consideration, which has significance for auxiliary decision making. The method of deriving urban form through machine learning can refer to design types that avoid high risks, which can be used as an important reference for urban planning and design in practical applications. In order to reduce the possibility of outbreaks of epidemic risk in urban space design, architects and researchers can make comparisons on the basis of the results derived from the distribution of epidemic hotspots, as well as adjust the design of urban textures, such as building density, roads, and green space layout.