1. Introduction
Atmospheric carbon dioxide (CO
2) levels are at their highest in recent history. Increasing CO
2 emissions have exacerbated the greenhouse effect and led to global warming, which in turn has spurred a series of environmental issues, such as sea level and temperature rise, increased incidence of extreme weather, and other potential hazards to global public health [
1,
2,
3]. According to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC), human activities have been the main drivers of global warming since the Industrial Revolution of the mid-20th century [
4]. As hubs for human social and economic activities, cities are reported to emit 71–76% of the total global energy-related carbon emissions, of which the building sector occupies approximately one-third [
5,
6]. For example, buildings account for more than 70% of the total energy consumption and CO
2 emissions in some major cities in the USA [
7], approximately 59% of electricity consumption in the European Union [
8], and over 60% of carbon emissions in Hong Kong, China [
9]. With continuous advancement of urbanization and improvement of living standards, CO
2 emissions from buildings are projected to increase, which warrants closer attention [
10,
11]. Therefore, emission reduction in the building sector is considered critical for effectively controlling the growth of CO
2 emissions from rapid urbanization. For emission reduction goals, it is essential to understand and assess CO
2 emission patterns in the building sector and investigate their spatiotemporal distributions, thus providing a foundation for the development of low-carbon cities.
For several decades, many domestic and international researchers have conducted valuable investigations on the assessment of CO
2 emissions from buildings. One of the most representative studies is the sustainable building assessment technical system proposed by the German Sustainable Building Association, which calculated the total life cycle carbon emission inventory of a building, including the production and construction, operation, maintenance, renewal, demolition, and reuse phases. In addition to the life-cycle-based method, some research institutions have proposed generic models for estimating energy consumption in the operational phase of existing buildings, such as the DeST model established by the School of Architecture, Tsinghua University [
12,
13], the Quick Energy Simulation Tool (eQuest) software [
14] and EnergyPlus [
15] developed by the US Department of Energy, and the CitySim operated by Swiss federal Institute of Technology in Lausanne [
16]. Although the above models and software can assess energy consumption and CO
2 emissions by simulating a building’s environment and equipment systems, most of them can only be applied to a single building. Such methods not only ignore the calculation of CO
2 emissions in buildings at the city scale and above, but they also fail to provide an in-depth spatial analysis of the urban carbon cycle and heat island effects. Gurney et al. [
17] used the eQuest model to estimate the energy consumption characteristics of buildings in Indianapolis and estimated the spatial and temporal distribution of carbon emissions for each building in the city. However, the building energy simulation tool requires complex field surveys and statistics, such as meteorological data and basic parameters of thermal disturbance, as model input parameters. In assessing the distributions of CO
2 emissions from urban buildings, such approaches, which require large-scale field investigations and statistical data, are costly and time-consuming, thereby limiting their application and efficiency in large regions.
The other widely used approach for CO
2 assessment at the citywide and larger scales is the top-down downscaling method that distributes emission sources from a large geographic region to smaller regions. Most existing downscaling methods allocate CO
2 based on spatial proxies such as night-time light imagery and population grid data. For example, on the basis of population data and administrative division data, Andres et al. [
18] established the spatial distribution data of CO
2 emissions, with a spatial resolution of 1°, from 1950 to 1990, and analyzed the growth characteristics of carbon dioxide emissions in different regions. Additionally, Oda and Aksyutov [
19] constructed an open-source data inventory of anthropogenic CO
2 (ODIAC) with a spatial resolution of 1 km, based on the relationship between night-time light images and the total carbon emissions of countries. Liu et al. [
20] used a linear relationship to calculate global daily residential CO
2 emissions during COVID-19 pandemic at the country scale. Zhao et al. [
21] downscaled building energy consumption carbon emissions at 1 km resolution by machine learning. These spatial proxies helped rationally allocate carbon emissions from buildings. For example, the Emissions Database for Global Atmospheric Research (EDGAR), developed by the European Commission Joint Research Centre (EU-JRC) and the Netherlands Environmental Assessment Agency (PBL), is one of the most representative inventories of carbon emissions, which utilizes energy and manufacturing facilities locations, road networks, the density of human and animal population, and a number of other spatial proxies. EDGAR provides spatially gridded data of global carbon emissions from buildings and other sectors of agricultural, transport, power, residential, industrial, and manufacturing [
22].
The global carbon grid includes global 0.1° × 0.1° CO
2 emission inventories in 2019 for the residential, power, industrial, transport, shipping, and aviation sectors with a calculation framework which integrates multiple data flows [
23,
24,
25]. However, the existing inventories of CO
2 emissions at large scales are generally operated within a grid unit (e.g., 1 km or 0.1° resolution), which may contain hundreds or thousands of buildings. Owing to the different types of human activities occurring in buildings (including studying, working, entertainment, and catering), the spatial distributions of CO
2 emissions from different buildings within the same grid might be significant differences. The grid-based inventories of CO
2 emissions cannot distinguish CO
2 emissions between different buildings, which may limit the assessment of the exact distributions of CO
2 emissions. Therefore, it is necessary to construct building-based spatial distributions of CO
2 emissions.
Further, the emission levels from individual buildings can vary over time due to changes in residential activities’ intensity. Specifically, the carbon emissions in a building differ significantly throughout the year as the different energy needs for heating/cooling are affected by seasonality [
20]. However, most current analysis methods and datasets of CO
2 emissions from individual buildings are based on a yearly basis or have a time lag of at least one year, making it difficult to analyze the temporal dynamics of CO
2 emissions accurately. To address this issue, some studies have attempted to establish reliable near-real-time data on carbon emissions. For example, EDGAR provides spatial maps of global carbon emissions with yearly, monthly, and hourly data [
22]. In addition, Liu et al. [
26] estimated near-real-time global daily CO
2 emissions on a 0.1° grid from sectors including power generation, industry and cement production, ground transportation, and commercial or residential construction. Unfortunately, such studies on the temporal dynamics of CO
2 emissions have only focused on the emissions within grid units and ignored the detailed spatial characteristics of buildings. As the building-based CO
2 emissions in large regions have rarely been analyzed and discussed, mapping building-based CO
2 emissions with temporal dynamics quickly and accurately remains a challenge.
With the emergence of ambitious climate policies and mitigation efforts [
27,
28], a reliable dataset of high-resolution spatiotemporal distributions of CO
2 emissions from buildings in large regions is needed to inform legislation. Based on these considerations, we attempted to integrate night-time imagery and building data to construct a monthly inventory of building-based CO
2 emissions using a case study in England, United Kingdom (UK). We first collected several datasets, including building data, Visible Infrared Imaging Radiometer Suite (VIIRS) night-time light imagery, and building-sector CO
2 emissions derived from EDGAR. We then built linear regression models for estimating the relationship between CO
2 emissions and two factors (building volume and night-time light) at the county level in each month. We selected the model with the best performance to help construct monthly CO
2 emission inventories for more than 11 million individual buildings in England. This study aimed to inform the development of a low-carbon city and promote sustainable development by analyzing the detailed spatiotemporal distribution of building-based CO
2 emissions in England.
3. Results and Discussion
3.1. Comparison of Model Performance
The statistical performance indicators for the three models are listed in
Table 1. For Model 1, the average values of
R2 and adjusted
R2 in each month were 0.862 and 0.859, respectively, while the MRE was approximately 34% and RMSE ranged from 24,000 to 90,000 tons. For Model 2, the average value of
R2 in each month was 0.858, and the average value of the adjusted
R2 was 0.855, while the MRE ranged from 46.62% to 50.34%. For Model 3, the values of
R2 and adjusted
R2 for each month were above 0.89, and
R2 reached the highest value of 0.911 in July, while the MRE for each month was below 26%. The RMSE of the model was the lowest of the three models, ranging from 20,000 to 75,000 tons. Evidently, Model 3 performed best among the three models, which indicates that both building volume and night-time lights are significant factors in estimating the CO
2 emissions from buildings, and considering both factors can greatly improve model performance.
To verify the effectiveness of the three models in estimating CO
2 emissions from buildings, we constructed scatterplots of county-scale true and estimated CO
2 emissions for regression Models 1–3, as shown in
Figure 6. Each model type was tested in January, April, July and October. The horizontal axis represents the actual CO
2 emissions from buildings in each county counted from the EDGAR grid map, and the vertical axis represents the CO
2 emissions from buildings estimated by the three models. For Model 1, the scatter points are more evenly distributed on both sides of the trend line, though there was considerable under- or over-estimation of CO
2 emissions in a few counties. For Model 2, most of the scatter points are below the one-to-one diagonal, and the RMSE of Model 2 was also the largest among the three models, suggesting that CO
2 emissions were severely underestimated. In Model 3, the points are adjacent to the one-to-one diagonal, which is the best fit of the three models. The
R2 and RMSE values further verify that the simulated values in Model 3 were closest to the predicted values. In July, the
R2 of Model 3 was enhanced from 0.862 to 0.911 when compared to the Model 1, and from 0.857 to 0.911 when compared to the Model 2, indicating that this model has the best fitting accuracy and highest feasibility. Moreover, the results also indicate that Model 3 maintains a degree of stability in estimating results in different months, with the value of
R2 varying only between 0.89 and 0.91.
Compared with the results from Model 1, which was constructed based only on building volume data, or Model 2, which was based only on night-time light as independent variables, regression models (Model 3) taking both explanatory variables into consideration together exhibit better performance and more stability, implying that both building volume and night-time lights are important factors for estimating CO2 emissions from buildings, and incorporating both variables can significantly improve the explanatory capacity of the model. Therefore, Model 3 is the most reliable and reasonable model for mapping the spatial distributions of building-based CO2 emissions. Subsequently, according to the regression model coefficients, Model 3 was selected for the estimation of the monthly CO2 emissions of each building in England.
To further evaluate the results, a multicollinearity diagnosis for Model 3 was necessary to verify the multicollinearity problem at the county scale.
Table 2 presents the multicollinearity test results for each variable. The VIF for two variables in each month is smaller than 10, demonstrating that weak correlations exist among these variables, and it is feasible to establish regression Model 3 with building data and nightlight data.
The monthly coefficients of Model 3 and the evaluation results are presented in
Table 2. For the model coefficients of each month, both
and
show significant seasonal variation, suggesting that the changes in night-time lights and building data in winter have a greater impact on CO
2 emissions than those in summer. The f-test significance results for the models in each month were 0, indicating that all models have a statistically significant predictive capability for CO
2 emissions from buildings. For the
t-test statistics, the significance values for night-time light data and building data in each month were less than 0.01, meaning that both variables can significantly affect the estimation of CO
2 emissions.
Thereafter, based on the coefficients obtained from Model 3, we can disaggregate the CO
2 emission data from the EDGAR grid map for more than 11.86 million buildings in England from January to December by using Formula (13):
where
is the CO
2 emissions from building
j in month
i;
and are the model coefficients in month
i;
is the volume of building
j; and
is the night-time light values of building
j in month
i.
3.2. National-Scale CO2 Emissions Analysis
Based on the models and coefficients in
Table 2, we successfully downscaled the CO
2 emissions from the EDGAR grids to individual buildings by coupling the building data and night-time light data. The spatiotemporal variations in CO
2 emission data from buildings at the national scale are presented in
Figure 7. CO
2 emissions from buildings in England showed obvious seasonality, being significantly higher in winter than in summer. EDGAR grid maps showed that, in winter, there were many areas scattered with red buildings, where the CO
2 emissions from a single building exceeded 70 tons a month, whereas in summer, most urban centers are only scattered with orange buildings, where the CO
2 emissions were less than 70 tons a month. According to one report [
54], the average temperature in England in January is 4–7 °C, and the average temperature in July is 13–17 °C. The fact that CO
2 emissions from buildings are significantly higher in winter than in summer may be due to winter heating, which causes more energy consumption.
From the perspective of spatial distribution, there is obvious spatial heterogeneity in CO
2 emissions from buildings. As seen in
Figure 7, most buildings are rendered green, which indicates that almost all the buildings emit less than 1 ton of CO
2 per month throughout the year. Only some counties, such as Greater London, Greater Manchester, and the West Midlands, contain red buildings whose monthly CO
2 emissions are much higher. Especially in the CBD areas, owing to frequent socio-economic activities, higher populations, higher density of buildings, larger building volumes, and other reasons, CO
2 emissions from buildings are significantly higher than those from other areas. Therefore, the distribution of CO
2 emissions from buildings are seemingly affected by the distribution of large cities. Furthermore, the spatial distribution of CO
2 emissions from buildings has a certain continuity, as buildings with high CO
2 emissions are always clustered and the CO
2 emissions of adjacent buildings have comparatively small differences. Overall, CO
2 emissions showed a decreasing trend from city centers to the surrounding areas.
We further adopted several statistical metrics, including the monthly maximum CO
2 emissions (Max CO
2), the monthly average CO
2 emissions (Ave CO
2), and the ratio of CO
2 emissions per volume (CO
2/Vol), to quantitatively evaluate the overall CO
2 emissions from buildings at national scale (
Table 3). For the specific implication of each metric, max CO
2 refers to the maximum CO
2 emissions from individual buildings during a month. Ave CO
2 is the average CO
2 emission from individual buildings during a month, which can reflect the overall situation every month. Another crucial measurement, CO
2/Vol, refers to the monthly CO
2 emissions produced per unit volume from buildings, which also indicates the CO
2 emission efficiency of buildings in different regions during different months.
The results show that CO2 emissions from buildings vary significantly between different buildings and by month. A single building can emit up to 3682 tons of CO2 in January and 894 tons in August (more than four times the difference), which exceeds the average CO2 emissions by nearly 4000 times. For CO2/Vol, the ratio of CO2 emissions per volume is 0.152 kg/m3 in January, which is 3.7 times higher than the ratios in July and August.
3.3. County-Scale CO2 Emissions Analysis
We obtained the maximum and average CO
2 emissions from buildings for each county in England (
Figure 8). We found that both maximum and average CO
2 emissions from buildings varied significantly in different counties and months. Clear “V” shapes in the boxplot indicate that the CO
2 emissions in nearly all counties decreased from January to July and increased from August to December. In particular, the discrepancy in CO
2 emissions between different counties was much larger in winter than in summer. It should be noted that the discrepancy in the maximum CO
2 emissions in all months was larger than the discrepancy in the average CO
2 emissions, which indicates that some buildings have extremely high CO
2 emissions.
We selected seven counties, and the corresponding statistical bar graphs were drawn from the ratio of CO
2 emissions per volume (CO
2/Vol) for each selected county, as shown in
Figure 9. The monthly changes in the CO
2/Vol of buildings in each selected county have the same “V-shaped” trend, representing slight declines from January to July, reaching the lowest value in July and August, and subsequently increasing from September to December. In addition, the absolute ratio of CO
2/Vol varied greatly among the different regions. For instance, as a typical metropolis, the CO
2/Vol of Greater London County was relatively high, at 0.206 kg/m
3 in January and 0.055 kg/m
3 in July. In Northumberland, in northern England, the CO
2/Vol was 0.081 kg/m
3 in January and 0.021 kg/m
3 in July. The CO
2/Vol of Greater London was 2.5 times higher than that in Northumberland, showing obvious regional differences.
Based on the CO
2 emissions from each county, we plotted a Lorenz curve between CO
2 emissions from buildings and the number of buildings in the whole England in January, as shown in
Figure 10. The horizontal axis represents the cumulative share of numbers of buildings, and the vertical axis represents the cumulative share of CO
2 emissions from buildings. The blue dotted line is the line of equality, which means the CO
2 emissions distribution is perfectly equal. The red line is the Lorenz curve; the greater the curvature, the more unequal the distribution of building CO
2 emissions. As the Lorenz curve shows, approximately 70% of the buildings account for 50% of the total CO
2 emissions. The area with the highest CO
2 emissions is Greater London, where the last 5% of buildings account for 15% of the country’s CO
2 emissions. Based on the Formulas (11) and (12), the environmental Gini coefficient can also be calculated. The range of the Gini coefficient is from 0 to 1, and a larger Gini coefficient means a higher degree of inequality. As a watershed, 0.4 is usually considered for whether the inequality level is too high [
55]. The Gini coefficient in England in January is 0.3479, which is less than 0.4, indicating that the disparity of CO
2 emissions allocation is reasonable.
3.4. Building-Scale CO2 Emissions Analysis of Typical City
To gain a better understanding and explore the CO
2 emission level of buildings in England, we selected urban centers of four counties located in different parts of England in January and July for further analysis.
Figure 11 shows CO
2 emission distribution maps at three different scales, including the EDGAR grid maps of four selected counties, building-based CO
2 emissions maps within a grid cell, and detailed CO
2 emission maps in three dimensions.
(1) London: Located in the southeast of England, London is the capital of the UK, as well as the political, cultural, and financial center [
29]. London has one of the most developed city economies, most prosperous businesses, and highest living standard in the world.
Figure 11a shows that the CO
2 emissions from buildings in almost all of the Greater London area were very high in January. The detailed map shows the CO
2 emissions from buildings in the commercial center near the Thames River, which has a large number of buildings and high building density. A large number of building CO
2 emissions decreased significantly in July, with most buildings falling from medium or high to relatively low levels. However, many large buildings still had very high emission levels in some areas of the north shore in July, reaching more than 250 tons of CO
2 per month. These high-carbon-emitting buildings include business centers, museums, and hotels that employ more electrical equipment, such as lifts and lights, leading to higher energy consumption.
(2) Manchester: Located in northwest England, Manchester is one of the largest metropolitan areas and one of the most important industrial centers in the UK. As shown in
Figure 11b, several buildings in the central urban area had extremely high CO
2 emissions in January, whereas most buildings in the city center emit less than 250 tons of CO
2 per month. In July, CO
2 emissions fell to moderate or low levels in almost all buildings, and the number of buildings with very high carbon emissions decreased. Compared with London, Manchester had significantly lower CO
2 emissions in both January and July.
(3) Bristol: Bristol is located in Avon County, a coastal area in southwestern England. Bristol is the largest city in southwest England and houses an important commercial port and space center. As shown in
Figure 11c, few buildings in the central city had extremely high CO
2 emissions in January, with the rest of the buildings in the central city having medium or low levels, showing an obvious decreasing trend from the central city to the surrounding suburbs. In July, CO
2 emissions fell to moderate or low levels in almost all buildings, except for a few that maintained high levels. Overall, CO
2 emissions from buildings in Bristol were slightly lower than those in Greater London.
(4) Cambridge: Cambridge is in eastern England. Compared to other cities, Cambridge is a nonmetropolitan county with lower socioeconomic activity. In the EDGAR gridded map, depicted in
Figure 11d, Cambridge had a low CO
2 footprint across the region in January and July. As shown in the detailed view, the downtown area of Cambridge had a lower number and density of buildings and a smaller base area for each building. However, although CO
2 emissions from buildings at Cambridge are generally low, there are still a few buildings that emit more than 250 tons of CO
2 in January and July.
A comparative analysis of these four cities shows that there is a significant seasonal variation in CO
2 emissions from buildings, with almost every building emitting significantly more CO
2 in winter than in summer. Additionally, from the perspective of spatial distribution, CO
2 emissions from buildings have an overall trend of decay from city centers to suburban areas, which means that buildings in city centers tend to have higher CO
2 emissions. Meanwhile, there is also spatial heterogeneity as indicated by the CO
2 emissions of some adjacent buildings, which may vary by up to a thousand times owing to large differences in building spatial structures. In addition, a comparison of the four counties in
Figure 11 shows that the number of buildings in different grids varies widely, suggesting that CO
2 emissions of the grid map are strongly related to the number of buildings in the grid and that the grid-based CO
2 inventories are not sufficient to show the detailed spatial–temporal distributions of CO
2 emissions for all buildings as grids with low CO
2 emissions possibly contain a few high CO
2-emitting buildings.
3.5. Policy Implications for CO2 Reduction
Reducing CO2 emissions from buildings with the requirement of green and low-carbon transformation of the economy and society is of great significance. The compilation of a CO2 inventory from the building sector is required to understand emission situations, establish emission baselines, verify emission trajectories, and develop efficient, feasible and viable mitigation options. This study developed linear regression models to calculate monthly building-based CO2 emissions. Unlike previous life-cycle-based methods or other approaches that require time-consuming field investigations and/or are based on a set of complex parameters, the mathematical method promoted in this study has been proven to be an effective way to quickly obtain building-based CO2 emissions over a large region by integrating only three accessible datasets, that is, gridded CO2 emissions from buildings, building volume data, and nightlight imagery. Therefore, this method can easily spread its use and may act as a highly efficient way to estimate the spatiotemporal distributions of building-based CO2 emissions when governments create low-carbon cities and sustainable development policies. More importantly, the spatial distributions of building-based CO2 emissions can aid in accurately identifying buildings with high CO2 emissions, which can offer useful support when prioritizing the priority decision of carbon reduction at a specific location.
However, the heating or cooling demands of residential and commercial buildings due to weather changes are the main reason for the large seasonal variation in CO
2 emissions in the building sector. According to a report by the Met Office, the average temperature in the UK ranges from 4 to 7 °C in January and 13 to 17 °C in July. The rapid rise in CO
2 emissions from buildings in England in winter is caused by extensive space heating, while a cooler climate with less cooling in summer results in lower CO
2 emissions. Some studies have recognized that improving heating and cooling systems is an effective way to reduce CO
2 emissions, including improving energy efficiency, increasing electrification, and creating cleaner electric grids [
56]. Therefore, the fast and time-efficient CO
2 emission estimating method with higher temporal resolution proposed in this paper can serve as a monitoring tool to provide references for policymakers in formulating and revising “decarbonization” strategy policies promptly.
3.6. Limitations
Further investigations are necessary to address the limitations of the present study. First, a thorough field survey is needed to obtain the actual CO2 emissions for buildings over a place, so as to provide true measured data to assess the accuracy of our estimated CO2 emissions from buildings. Second, we only considered two building factors: building volume and night-time light. Other factors, such as population distributions and energy consumption demands, might have significant impacts on CO2 emissions and should thus be considered. Third, fine-scale CO2 emission data from different types of buildings (e.g., building function, design, orientation) can provide valuable references for decision makers to develop different carbon reduction policies. However, due to the lack of a breakdown of buildings, the present study failed to consider the differences in building types when modeling building-based CO2 emissions in the studied region, which may introduce uncertainties into our results. Therefore, future researchers should consider the types of building when modeling CO2 emissions from buildings in order to identify super emitter categories.
It should be noted that the monthly CO2 emissions data used in our study is the EDGAR v5.0 dataset published in November 2019, which provides monthly CO2 emissions data from 1975 to 2015. It was the latest version available from the EU-JRC when we did the experiment. Recently, the latest version of EDGAR v6.0 was published, which provides CO2 emissions data for building sector from 1970 to 2018. Collecting building data and night-time light data for 2018 and using our proposed methodology to calculate monthly CO2 emission levels for individual buildings in 2018 and other years would be an easy and predictable work in the future.