Applied Energy 131 (2014) 297–306

Contents lists available at ScienceDirect

Applied Energy

journal homepage: www.elsevier .com/locate /apenergy

Spatial effects of carbon dioxide emissions from residential energyconsumption: A county-level study using enhanced nocturnal lighting

http://dx.doi.org/10.1016/j.apenergy.2014.06.0360306-2619/� 2014 Elsevier Ltd. All rights reserved.

⇑ Corresponding author at: Institute of Natural Resources and EnvironmentalScience/Henan Collaborative Innovation Center for Coordinated Developments inCentral China Economic Zone, Henan University, Kaifeng 475004, China.

E-mail addresses: cluheli@gmail.com (H. Lu), kf_guif@163.com (G. Liu).

Heli Lu a,b,⇑, Guifang Liu a

a Institute of Natural Resources and Environmental Science/Henan Collaborative Innovation Center for Coordinated Developments in Central China Economic Zone, HenanUniversity, Kaifeng 475004, Chinab United Nations University, Institute of Advanced Studies, Yokohama 220-8502, Japan

h i g h l i g h t s

�We evaluate carbon dioxide emissions from residential energy consumption (REC) at county-level.� A stepwise methodological procedure is conducted using satellite imagery of nighttime lights.� We find a high degree of county-level clustering in the distribution of emissions per capita.� High-emission counties tend to be surrounded by counties with relatively low per capita GDP levels.� We stress the need for consideration of other factors in determining emission patterns.

a r t i c l e i n f o

Article history:Received 5 February 2014Received in revised form 17 April 2014Accepted 17 June 2014Available online 10 July 2014

Keywords:Carbon dioxide emissionsResidential energy consumptionNighttime lightsSpatial effects

a b s t r a c t

As the world’s largest developing country and greenhouse gas emitter, China’s residential energyconsumption (REC) is now responsible for over 11% of the country’s total energy consumption. In thispaper, we present a novel method that utilizes spatially distributed information from the Defense Mete-orological Satellite Program’s Operational Linescan System (DMSP–OLS) and human activity index (HAI)to test the hypothesis that counties with similar carbon dioxide emissions from REC are more spatiallyclustered than would be expected by chance. Our results revealed a high degree of county-level clusteringin the distribution of emissions per capita. However, further analysis showed that high-emission countiestended to be surrounded by counties with relatively low per capita GDP levels. Therefore, our resultscontrasted with other evidence that REC emissions were closely related to GDP levels. Accordingly, westress the need for the consideration of other factors in determining emission patterns, such as residentialconsumption patterns (e.g., consumer choices, behavior, knowledge, and information diffusion).

� 2014 Elsevier Ltd. All rights reserved.

1. Introduction

The Intergovernmental Panel on Climate Change (IPCC) hasreported that global surface temperatures have increased by0.74 ± 0.18 �C (1.33 ± 0.32 �F) during the 20th century. This is likelythe result of increasing concentrations of greenhouse gases (GHGs)[1]. Although human activities, including the burning of fossil fuelsand deforestation, were only responsible for 3% of carbon dioxide(CO2) emissions worldwide in the 1990s, the increase in emissionswas large enough to exceed the absorptive capacity of naturalprocesses (e.g., photosynthesis). Energy use for power generation

in the industrial [2,3], residential [4,5] and transportation sectors[6,7] has been the largest contributor of GHG emissions. Inparticular, approximately 10–35% of national energy consumptionwas from the residential sector worldwide in the late 1990s and2000s [8], and this figure will likely increase as the averageperson’s income, standard of living, and associated access to homeappliances, housing and personal transportation increase. Accord-ingly, great attention has recently been placed on the residentialsector’s role in generating GHGs [9].

In China, the residential sector is now the second greatestenergy consumer behind the industrial sector; total energyconsumption in the residential sector grew by 8.43% since 2008compared to an average increase of 2.58%. The increase in energyconsumption by this sector has been driven by three major factors.First, the Chinese government has refocused its economy towardsdomestic consumption, resulting in a household consumption

298 H. Lu, G. Liu / Applied Energy 131 (2014) 297–306

expenditure that was 38.2% of the Gross Domestic Product (GDP)[10]. If the goal of shifting the Chinese economy towardsconsumption-led growth in the 11th Five Year Plan is met, it is veryprobable that residential energy consumption (REC) will continueto rise in the future [11]. Secondly, the lifestyles and consumerpreferences of Chinese residents have changed as the country con-tinues to develop, with greater individual access to high qualityfood, comfortable living environments, health care, personalhygiene products and higher education. Finally, the annual percapita energy consumption of urban Chinese residents is 3.5 timesthat of rural residents, and, with the urban population expected togrow by 20 million per year, the rapid growth of REC is likely tocontinue.

Given recent increases in REC, estimating associated CO2 emis-sions from this source in a spatially explicit manner is critical forcombating processes like global climate change. Estimating emis-sions from REC is usually dependent on statistical data [12], mak-ing it difficult to produce a spatially realistic representation ofemissions. However, because artificial lighting is a unique indicatorof residential activity, datasets derived from satellite imagery ofnighttime lights have been used to map phenomena which wouldotherwise be difficult to map through ground surveys. There havebeen many studies examining the use of these datasets for deter-mining spatiotemporal dimensions of socio-economic factors,including GDP, urban sprawl, impervious surfaces, and ex-urbandevelopment. More recently, these datasets of nighttime lightshave been used as proxies for estimating CO2 emissions. The firstglobal map of CO2 emissions had a spatial resolution of 1 degreeand was developed by combining the lighted area of a city (usingimagery of nighttime lights acquired between October 1994 andMarch 1995) with country-level ancillary statistical information[13]. Oda and Maksyutov [14] created a high resolution globalinventory of annual CO2 emissions for the years 1980–2007 bycombining a worldwide point source database with satellite obser-vations of global nighttime lights distribution. Finally, Ghosh et al.[15] developed a method of mapping spatially distributed CO2

emissions from the burning of fossil fuels (excluding electric powerutilities) at 30 arc-seconds (approximately 1 km2 resolution) usingregression models of nighttime lights images.

In this paper, we present a new approach that uses the humanactivities index (HAI) as auxiliary data to correct saturatednighttime lights and to fill values in areas lacking nighttime lights.Using our methodology, satellite images of nighttime lights mayserve as a useful proxy for the distribution of CO2 emissions fromREC. Finally we used the emissions distribution to test the hypoth-esis that counties with similar CO2 emissions from REC are morespatially clustered than would be expected by chance.

Fig. 1. Methodological framework used in this study.

2. Data and methodology

We used images of nighttime lights collected by the U.S. AirForce Defense Meteorological Satellite Program’s (DMSP) Opera-tional Linescan System (OLS), which has been cited as a remarkableexample of a global earth observing satellite sensor for detectinghuman activity [16]. This low orbiting satellite uses the visible/near infrared waveband (0.4–1.1 lm) for detecting lights and thethermal infrared (10.5–12.6 lm) band to filter cloud cover [17].The satellite typically makes passes a study area between 8:30p.m. and 9:30 p.m. local time, and annual global composites oftemporally stable nighttime lights have been produced by theNational Geophysical Data Center (NGDC) covering the period from1992 to 2008 [18]. There are three versions of the data available fordownload, including (1) ‘‘raw,’’ (2) ‘‘stable lights’’, and (3) ‘‘cali-brated’’ versions [19]. We used the stable lights product (spatialresolution: 1 km2), which removes clouds, gas flares, lightning,

and other ephemeral and extraneous signals using the proceduredescribed by Elvidge et al. [20]. Images used captured nighttimelights over mainland China for the year 2000. Using these images,we conducted a new stepwise methodological procedure (Fig. 1) tomap CO2 emissions from REC.

2.1. Calculation of CO2 emissions from REC

To calculate CO2 emissions from REC, we used residentialenergy matrix tables for 12 fuel types from China’s Energy Statisti-cal Yearbook for the year 2000 [21] to estimate CO2 emissions inthat year in rural and urban areas separately. Following IPCCguidelines [22], we used the following equation to calculate CO2

emissions from REC:

ECO2 ¼X

j

Aj � ALCj � Cj � Oj � 44=12 ð1Þ

where ECO2 is the total CO2 emissions from energy consumption(tons), Aj is the amount of fuel consumption j (tons or cu m (cubicmeter) for gas), ALCj is the average low calorific value of fuel j (kJ/kg or kJ/cu m), Cj is the carbon emission factor of fuel j (kg C/GJ),and Oj is the carbon oxidation rate of fuel j. The average low calorificvalues, carbon emission factors, and carbon oxidation rates for eachfuel j are listed in Table 1. The coefficient of emission conversionfrom heat, a secondary energy source, was also used in our study.

2.2. Threshold method

We identified urban and rural areas based on nighttime lightcoverage for further construction of the HAI regression model.Satellite imagery has been widely used to map urban sprawl,although confusion exists around the use of empirical thresholdsto delineate city boundaries. Imhoff et al. [23] has suggested thata threshold of 89% could remove ephemeral light sources and the‘‘blooming’’ of light onto water bodies adjacent to cities while stillleaving the dense urban core intact. However, Small et al. [24] hasargued that a single threshold for all cities could result in an

Table 1Average low calorific value, carbon emission factor and carbon oxidation rate.

Fuel type Carbonemission factor(Kg C/GJ)

Averagelow calorificvalue (kJ/kg)or (kJ/cu m)

Carbonoxidation rate

Raw coal 25.8 20,908 1Cleaned coal 26.7 26,344 1Coke 29.2 28,435 1Coke oven gas 12.1 17,353 1Gasoline 18.9 43,070 1Kerosene 19.6 43,070 1Diesel oil 20.2 42,652 1Fuel oil 21.1 41,816 1Liquefied petroleum gas 17.2 50,178 1Refinery gas 18.2 46,055 1Natural gas 15.3 38,931 1Heat 9.46

H. Lu, G. Liu / Applied Energy 131 (2014) 297–306 299

inconsistent relationship between the lighted area and the actualboundaries of urban areas. In this study, we used a dynamicthreshold from Chunyang et al. [25] in order to obtain urban spatialpattern characteristics for each province in mainland China basedon statistical data for urban land area (according to China’s Minis-try of Land and Resources). A suitable threshold was found at thepoint of least error between the accumulation of light area andactual statistical data using a two-step calculation. In the first step,we calculated the total urban area as:

Nighttimearea ¼XDNmin

DNi¼DNmax

DNiarea ð2Þ

where DNi is one step value from DNmax (e.g., 63) down to DNmin

(e.g., 1), given that urban areas are more well-lit than rural areas;and DNiarea is the urban land area derived from the pixels in thenighttime lights imagery with the value DNi. In step two, we com-pared Nighttimearea at DNi to statistical data in order to determinethe appropriate threshold according to the following relationship:

jNighttimeareaDN i� Landuseareaj � e ð3Þ

where NighttimeareaDN iis the urban area based on nighttime lights at

DNi; Landusearea is the actual urban area based on statistical data;and e is the minimum difference between Nighttimearea andLandusearea for all step values from DNmax to DNmin. Appropriatethresholds for delineating urban boundaries were determined atthe point where Eq. (3) was satisfied.

2.3. Construction of HAI

As discussed previously, nighttime lights have been used as aproxy for measuring human impacts on the environment and othersocio-economic phenomena [26]. The magnitude of humanimpacts can also be captured with HAI. Thus, it could be assumedthat these two indicators are interwoven such that the relationshipcould have some exact functional representation. In this study, weused HAI to account for REC in areas lacking nighttime lights orwhere the saturation effects of nighttime lights were high. Here,HAI analyses were based on the assumption that growth in humanactivity could be attributed to growth in economic activity (asmeasured by residential income index (II)) or to the humanhabitation environment (as quantified by the human settlementenvironment index (HSEI)).

We calculated HAI as the geometric mean of II and HSEIaccording to the following equation:

HAIi;j ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiIIi;j �HSEIi;j

qð4Þ

where HAIi,j is the human activity index, IIi,j is the residential incomeindex, and HSEIi,j is the human settlement environment index, all atgrid cell i in province j. The equation defining II has been promul-gated by the United Nations Development Programme [27] and iscalculated by the natural logarithm of GDP per grid cell [28]:

II ¼ lnðGDPÞ � lnðminðGDPÞÞlnðmaxðGDPÞÞ � lnðmin GDPð ÞÞ ð5Þ

The intensity and magnitude of HSEI in this study was derivedfrom an HSEI model with a large range of parameters, includingsurface relief, climatic condition and land cover type, as developedby Feng et al. [29]. We accomplished this using the following steps:

Step 1. The complexity of surface relief often determines thelocalization of human activities in a particular area or limits ofutilization of an area for a given activity. The surface relief index(SRI), which reflects the difference in regional terrain in a defineddistrict, could be calculated as:

SRI ¼ Ave Alt100þ ½MaxðAltÞ �MinðAltÞ� � 1� Flat Area=Area½ �f g=500 ð6Þ

where Ave_Alt is the average altitude (m); Alt is altitude (m);Flat_area is the area (km2) of a study region with a slope <5�; andArea is the total area of the study region. We defined the districtas a 5 � 5 cell square window around a grid cell.

Step 2. Among the bioclimatologic indices, thermo-higrometricindex (THI), which was used originally to determine discomfortdue to heat stress [30], was chosen to evaluate climatic effectson human activities. We used the following formula for determin-ing the THI index [31,32]:

THI ¼ T � ð0:55� 0:0055 � RHÞ � ðT � 14:5Þ ð7Þ

where T is air temperature measured in degrees Celsius, and RH isthe relative humidity of the air.

Step 3. The interaction between human and land cover affectsthe states and dynamics of human occupants. We used the follow-ing formula to calculate land cover index (LCI):

LCI ¼ NDVI � LTi ð8Þ

where NDVI is the normalized difference vegetation index, and LTi isthe coefficient of various land cover types. Finally, based on thequantitative analysis of SRI, THI, and LCI, we used the followingformula to calculate HSEI:

HSEI ¼ a � SRIþ b � THIþ c � LCI ð9Þ

where a, b, and c are the weight of SRI, THI and LCI, respectively, asderived from Feng et al. [29].

To determine SRI, THI, and LCI, we used a variety of additionaldatasets. For SRI, we used DEMs (spatial resolution: 1 km) fromthe U.S. Geological Survey (USGS). To calculate THI, we used datafrom 578 meteorological stations within the Chinese Meteorologi-cal Network [33] to calculate average annual temperature andrelative humidity from 1961 to 2006. The climatic distributionmaps were generated via Kriging spatial interpolation methods inARCGIS. For LCI, we used NDVI datasets derived from the averagemaximum annual value of SPOT-VEGETATION data from 1998 to2006 [34]. We also obtained a land cover map of China for the year2000 from the Chinese Earth System Science Data Sharing Net-work. For the coefficients for land cover type, we utilized the tech-nical criterion value for the evaluation of eco-environmental statusas recommended by the State Environmental Protection Adminis-tration of China [35]. Finally, all spatial data were re-sampled tothe same resolution as DMSP–OLS images and converted into theAlbers Conical Equal Area projection before conducting spatialcomputation.

300 H. Lu, G. Liu / Applied Energy 131 (2014) 297–306

2.4. Moran’s I statistics

We analyzed spatial clustering patterns in emissions data usingglobal and local Moran’s I. To determine whether emissions werespatially correlated, we used global Moran’s I, which is a statisticalmeasurement of spatial autocorrelation between county-levelemissions [36–38] according to the equation:

I ¼NP

i

Pjwijðxi � �xÞðxj � �xÞ

ðP

i

PjwijÞ

Piðxi � �xÞ2

ð10Þ

where N is the number of counties; wij is the spatial weights matrix;xi and xj are county-level emissions for counties i and j, respectively;and �x is the mean value of county-level emissions. Moran’s I variesbetween �1 and 1, where a value near 1 indicates that similarattributes are clustered, a value near �1 indicates that dissimilarattributes are clustered, and a value near 0 is indicative of a randompattern lacking spatial association. We also tested the statistical sig-nificance of Moran’s I by (1) randomizing the data over space andcalculating a single value for the Moran’s I statistic, (2) obtainingan empirical distribution of Moran’s I under the null hypothesis ofno spatial autocorrelation via 999 randomization iterations, and(3) comparing the distribution from the previous step with the Mor-an’s I calculated with the original, non-randomized data.

To further explore the spatial pattern of county-level emissions,we also investigated where emission concentrations occurred andwhether high- or low-emissions values were clustered. We did thisusing local Moran’s I [39], which captures local indicators of spatialassociation (LISA), local pockets of instability, or local clusters. Wecalculated local Moran’s I according to the following equation:

Ii ¼ðxi � �xÞ

s2x

Xj

wijðxj � �xÞ� �

ð11Þ

where x is county-level emissions. This statistic describes fourbroad patterns of local spatial autocorrelation. A spatial clustercan either have a high value (for emissions) surrounded by similarlyhigh values (positive Ii), low values surrounded by low values(positive Ii), or high values surrounded by low values (and viceversa; negative Ii) [40].

Finally, because it is generally believed that economic growthand emissions are highly correlated [41,42], we further analyzedspatial autocorrelation between emissions clusters and GDPthrough a bivariate Moran scatter [38].

3. Results

3.1. Urban–rural spatial pattern

According to the nighttime light imagery, we mapped distinctregions for urban areas with nighttime lights as well as rural areaswith and without nighttime lights. Fig. 2 shows original DMSP–OLSnighttime lights satellite imagery (a), from which patterns (b) forurban areas with nighttime lights, rural areas with nighttime lightsand rural areas without nighttime lights were derived. Here, weonly present the examples for Beijing and Hebei while this analysiswas also conducted for other provinces. Table 2 shows 18provinces with minimal relative error (<5%) in actual urban areacompared to the urban area derived from DMSP–OLS nighttimelights.

3.2. Regression model

Cluster analysis is an exploratory data analysis tool used to sortdifferent objects into groups such that objects in the same grouphave maximal associations and those in different groups haveminimal associations. Based on the Euclidean distance measure

of cluster analysis, we classified provinces into two groups accord-ing to their similarities in the percentage of urban area and averagenighttime lights (Fig. 3). Group A included three mega-cities(Beijing, Shanghai and Tianjin) and Group B included other prov-inces. A second-order polynomial regression model (Model 1)and an exponential regression model (Model 2) best fit the datain Group A and Group B, respectively. The second-order polynomialregression model associated with Group A resulted in coefficientsof determination values (R2) of 0.82, 0.93 and 0.84 for the citiesof Tianjing, Beijing, and Shanghai, respectively (Table 3). Thismodel indicated a saturation effect for the nighttime lights in thesecities (Fig. 4). Because of this effect, the response of lights againstHAI reached a finite amount and remained at that level withoutexceeding it.

In Group B, an increase in HAI did not produce a saturationeffect. Therefore, a bounded functional form, such as a second-order polynomial mentioned above, was not suitable, and a moresatisfactory model for these data should be developed. The rela-tionship between HAI and nighttime lights was reliably expressedas an exponential function. Unlike a second-order polynomialfunction, the exponential function considered only positive values,and nighttime lights increased quickly as HAI increased (Fig. 4).The exponential regression model in Group B resulted in a goodfit of R2 values >0.70 for all provinces (Table 3).

Then we applied the regression parameters of Model 1 andModel 2 (Table 3) to HAI to produce an inversion of nighttime lightsin areas lacking lights. Here, we present the illustration for Beijingand Hebei in Fig. 5. The white area, which indicates no nighttimelights, was replaced by blue or red in right graphs. In particularthe nighttime light intercept for Model 1 was negative, implyingthat many anthropogenic activities in the mega-cities could notbe reflected by nighttime lights. As a result, to maximize the infor-mation from the inversion, the nighttime lights axis was shifted tothe bottom by the intercept distance in order to avoid negativevalues and reduce the ‘saturation effect’ in mega-cities. As withthe first graph of Beijing in Fig. 5, the change of the highest valuein the two graphs (a and b) was dramatic, shifting from 63 to 83.

3.3. Model validation

It is difficult to directly validate the relationship betweenenhanced nighttime lights and CO2 from REC at county-level sincelack of corresponding statistical data. To achieve the goal weadopted an indirect measurement. First we checked the relation-ship between CO2 from REC and main part of CO2 from REC(including natural gas and liquefied petroleum gas) in 153 countiesin 2000 listed in the China City Statistical Yearbook using derivedHAI. The result showed a relatively high linear correlation betweenCO2 from REC and main part of CO2 from REC (R2 = 0.74). Then afterchecking the relationship between enhanced nighttime lights andmain part of CO2 from REC we again determined a high linear cor-relation of R2 = 0.77. As a result we believed that this analysis sug-gested that emissions for these counties were in proportion to thenumber of enhanced nighttime lights. Given this relationship, wewere able to divide the provincial REC emissions calculated fromthe ‘China’s Energy Statistical Yearbook 2001’ by the total numberof enhanced nighttime lights, yielding emissions per light unit foreach province. They were subsequently aggregated at county-levelto estimate CO2 emissions from REC in China.

3.4. Spatial autocorrelation analysis

We plotted county-level emissions against spatial lag, whichcorresponded to the weighted average emissions (per capita andlogarithmic) for a county’s neighbors. In Fig. 6, counties with highemissions that were surrounded by neighbors with high emissions

Fig. 2. Urban–rural spatial pattern in Beijing and Hebei.

H. Lu, G. Liu / Applied Energy 131 (2014) 297–306 301

fell into quadrant HH, counties with low emissions that had neigh-bors with high emissions fell into quadrant LH, counties with lowemissions that were surrounded by counties with low emissionsfell into quadrant LL, and counties with high emissions that hadneighbors with low emissions fell into quadrant HL. We found asignificantly high spatial association in China’s county emissions(standardized log of emissions per capita), with a global Moran’sI value of 0.58 (p < 0.0001). Thus, counties with relatively high(low) emissions levels were located near other cities with high(low) emissions levels more often than would be expected due to

random chance. For example, counties in quadrant HH and quad-rant LL accounted for 46.8% and 37.0% of the total, respectively.

In addition, spatial autocorrelation analysis between standard-ized per capita CO2 emissions and Gross Domestic Product (GDP)at county-level was conducted. In Fig. 7, counties with high emis-sions that were surrounded by counties with high GDP fell intoquadrant HH, counties with low emissions that were surroundedby counties with low GDP fell into quadrant LL, counties withlow emissions that were surrounded by counties with high GDPfell into quadrant LH, and counties with high emissions that were

Table 2Accuracy assessment for urban areas derived from DMSP/OLS nighttime light satellite imagery compared to actual urban area.

Province name Total area of province (109 km2) Urban area – statistical (km2)a Urban area – satellite (km2)b Relative error (%)

Beijing 28.01 2272.60 2269.56 0.13Tianjin 19.37 2254.75 2285.62 �1.37Hebei 314.54 14435.26 15107.99 �4.66Shanxi 248.74 7236.02 7478.21 �3.35Liaoning 257.08 10940.71 11203.26 �2.40Jilin 363.83 8210.65 8015.23 2.38Heilongjiang 1005.63 11369.11 11359.85 0.08Shanghai 8.50 2083.30 2069.81 0.65Jiangsu 143.20 13575.63 13247.97 2.41Zhejiang 133.44 5863.14 6042.79 �3.06Fujian 150.45 4243.23 4161.70 1.92Shandong 237.08 18806.06 18494.77 1.66Henan 239.83 18538.14 18505.81 0.17Hubei 252.66 9573.93 9790.93 �2.27Guangdong 210.04 12380.66 12341.55 0.32Yunnan 466.19 5704.57 5680.43 0.42Shaanxi 650.09 6844.25 6667.15 2.59Xinjiang 19.37 8860.04 8768.07 1.04

a Statistical data for urban land area from China’s Ministry of Land and Resources.b Derived from DMSP–OLS nighttime lights.

Fig. 3. Clustering of mega-cities (triangles) and other provinces (diamonds) inChina.

Table 3Parameters in Model 1 and Model 2.

Name Model type Function R2

Beijing 1 Y = �0.0084x2 + 1.5533x � 21.591 0.82Tianjin 1 Y = �0.0039x2 + 0.7754x � 13.633 0.93Shanghai 1 Y = �0.0055x2 + 1.0506x � 9.549 0.84Shanxi 2 Y = 0.357e0.0373x 0.96Liaoning 2 Y = 0.9559e0.0259x 0.98Jilin 2 Y = 1.2075e0.0215x 0.93Heilongjiang 2 Y = 0.8673e0.0281x 0.75Hebei 2 Y = 0.6616e0.0268x 0.90Jiangsu 2 Y = 0.5354e0.0317x 0.79Zhejiang 2 Y = 0.7678e0.0305x 0.95Fujian 2 Y = 0.6223e0.0341x 0.95Shandong 2 Y = 0.1417e0.0428x 0.98Henan 2 Y = 0.2358e0.0339x 0.99Hubei 2 Y = 0.9737e0.0175x 0.95Guangdong 2 Y = 1.9636e0.0271x 0.96Yunnan 2 Y = 0.2904e0.0344x 0.90Shaanxi 2 Y = 1.0043e0.0205x 0.73Xinjiang 2 Y = 2.5838e0.0185x 0.74

Note: Model 1 indicates second-order polynomial regression model while Model 2indicates exponential regression model. X indicates HAI while Y indicates Night-lights in rural area.

Fig. 4. Model 1 (second-order polynomial regression model) and Model 2 (expo-nential regression model).

302 H. Lu, G. Liu / Applied Energy 131 (2014) 297–306

surrounded by counties with low GDP fell into quadrant HL.According to the bivariate Moran scatter for quadrant HH inFig. 6, about five times more counties fell into the High emissions

– Low GDP quadrant compared to the number that fell into theHigh emissions – High GDP quadrant. Thus, counties that produceda high volume of emissions tended to be surrounded by countieswith relatively low per capita GDP levels.

4. Discussion

In this study, we used a new methodological approach thatcombined DMSP–OLS satellite imagery of nighttime lights andthe HAI. The novel use of HAI served as an important proxy for

Fig. 5. Conversion of (a) original DMSP–OLS nighttime lights to (b) enhanced nighttime lights based on Model 1 and Model 2.

H. Lu, G. Liu / Applied Energy 131 (2014) 297–306 303

human activity in regions where nighttime lights were unavailable.It also demonstrated the role played by the integration of socio-economic and environmental factors, such as GDP, digital DEMs,temperature, humidity, NDVI and land cover types. We ultimatelyfound that the resulting index was statistically correlated with CO2

emissions from REC at county-level.County-level emissions from REC rarely occur in isolation. Our

study highlighted that a high degree of county clustering occurredin the distribution of emissions. The challenge in this study, how-ever, is to identify the true cause of the emission pattern. Based on

the present study’s analysis, it cannot be explained by a single, orperhaps even primary, factor like GDP. While some results pointtowards economic growth as a contributing factor [43,44], thereis no way to establish a causal relationship. Results from the spatialassociation within our study, however, found that energyrequirements, CO2 emissions, and economic development levelsare correlated and that the correlation is far from absolute.

Consequently, we conclude that there should be other determi-nants of REC, such as consumer choices, behavior, knowledge, andinformation diffusion. For example, there may be consumer

Fig. 6. Spatial autocorrelation in CO2 emissions at county-level.

Fig. 7. Spatial autocorrelation between standardized per capita CO2 emissions and Gross Domestic Product (GDP) at county-level.

304 H. Lu, G. Liu / Applied Energy 131 (2014) 297–306

choices, behavior, information, or knowledge that arises from otheractivities in neighboring counties that come into play in a region.Conversely, if households in a county change current consumerpatterns, the effect of this type of change may spread to adjacentareas by means of, for example, advertisements, societal norms,habits, or constraints. This information can offer clues concerningemission reduction potential. For example, a change towards agreener consumption pattern could have expansive effects onreducing the emissions of households and individuals.

5. Conclusions

In this paper, we discuss a new approach for using enhanceddatasets of nighttime lights to estimate CO2 emissions from RECat county-level. With recent advances in economic developmentand extremely high human population densities, China has becomethe largest emitter of GHGs worldwide, and a large volume of thoseemissions have come from REC. Although most estimates of CO2

emissions from REC are given by coarse-grained statistics [45,46],

H. Lu, G. Liu / Applied Energy 131 (2014) 297–306 305

fine-scale evaluations of emissions at a regional level are essentialfor policy makers to improve energy management, to reducebuilding energy use, and to develop standard solutions for energysavings. More specifically, a county-level distribution pattern foremissions is needed to more effectively monitor planning and pol-icy implementation for meeting emissions targets in the REC sectorin China. Our methodological approach would allow policy makersto track residents’ carbon footprints to meet this information need.We believe the methodology developed in the research particularlyhas great significance in developing countries due to unavailabilityand poor quality of their statistical data. Easily-accessed andreliable comparable data at smaller areas in these countries couldprovide an appropriate base from which compromises for meetingthe challenge of climate change could quickly be reached. Further-more with HAI-Nightlights model REC emissions could be fast,simple and easy to be updated based on annual updated DMSP–OLS satellite images by NGDC. As a result the study on RECemissions in 2000 as base year with which the values during otherperiods could be compared not only describe a basic emissioncharacteristic and intensity in China but also provide a good start-ing point for further change detection and future dynamic analysisof time series emission maps.

Our study also explored how county-level characteristics (e.g.,distribution of emissions from REC) might be linked to real eco-nomic activities. Understanding scale and geographic scope arecritical for understanding patterns of CO2 emissions in China[47,48]. Our results provide empirical evidence of generalspillovers and spatial homogeneity among countries in terms ofemissions per capita. However, since the interaction processes forGDP vary between counties, emission clustering could be the resultof factors other than the level of economic development. Thus,understanding these spatial correlations will require more precisemeasurements of both observed factors in residential incomes andcross-border flows associated with consumer patterns. This pro-vides a promising framework for the next stages of REC-basedresearch. For example, through the combination with aerial photo-graphs or high-resolution satellite images which could provideabundant shape structure and texture information of residentialbuildings, the spatial resolution on evaluation of REC emissionscould be efficiently improved. Additionally the change informationon REC sector in time series through annually updated emissionmap from DMSP–OLS data will have great significance in improv-ing the management ability and enhancing the working efficiencyof government on residential carbon emission reduction.

Acknowledgements

The authors thank the anonymous reviewers whose commentsand suggestions were very helpful in improving the quality of thispaper. The authors also thank the editor for helpful suggestions.This project is funded with support from National Natural ScienceFoundation of China under Project 41371525, National BasicResearch Program of China (973 Program) 2012CB955800(2012CB955804), China Postdoctoral Science Foundation fundedProject (2012M521390 and 2013T60696), Scientific ResearchFoundation for Returned Scholars 2013(693) and 2013B065.

References

[1] Solomon S, Qin D, Manning M, Chen Z, Marquis M, Averyt KB. Climate change2007: the physical science basis. New York: Cambridge University Press; 2007.

[2] Yang X, Teng F, Wang GH. Incorporating environmental co-benefits intoclimate policies: a regional study of the cement industry in China. Appl Energy2013;112:1446–53.

[3] Ammar Y, Joyce S, Norman R, Wang YD, Roskilly AP. Low grade thermal energysources and uses from the process industry in the UK. Appl Energy2012;89:3–20.

[4] Zhu D, Tao S, Wang R, Shen HZ, Huang Y, Shen GF, et al. Temporal and spatialtrends of residential energy consumption and air pollutant emissions in China.Appl Energy 2013;106:17–24.

[5] Yang L, Yan HY, Lam JC. Thermal comfort and building energy consumptionimplications – a review. Appl Energy 2014;115:164–73.

[6] Chung W, Zhou GH, Yeung IMH. A study of energy efficiency of transport sectorin China from 2003 to 2009. Appl Energy 2013;112:1066–77.

[7] Zamboni G, Malfettani S, André M, Carraro C, Marelli S, Capobianco M.Assessment of heavy-duty vehicle activities, fuel consumption and exhaustemissions in port areas. Appl Energy 2013;111:921–9.

[8] Lenzena M, Mette W, Claude C, Hitoshi H, Shonali P, Roberto S. A comparativemultivariate analysis of household energy requirements in Australia, Brazil,Denmark, India and Japan. Energy 2006;31(2–3):181–207.

[9] Kelly S, Shipworth M, Shipworth D, Gentry M, Wright A, Pollitt M, et al.Predicting the diversity of internal temperatures from the English residentialsector using panel methods. Appl Energy 2013;102:601–21.

[10] Chen S, Yoshino H, Levine MD, Li Z. Contrastive analyses on annual energyconsumption characteristics and the influence mechanism between new andold residential buildings in Shanghai, China, by the statistical methods. EnergyBuild 2009;41(12):1347–59.

[11] Zhou N, McNeil MA, Levine M. Energy for 500 million homes: drivers andoutlook for residential energy consumption in China. Berkeley,California: Environmental Energy Technologies Division, Ernest OrlandoLawrence Berkeley National Laboratory; 2010.

[12] Yu SW, Wei YM, Fan JL, Zhang X, Wang K. Exploring the regional characteristicsof inter-provincial CO2 emissions in China: an improved fuzzy clusteringanalysis based on particle swarm optimization. Appl Energy 2012;92:552–62.

[13] Doll CNH, Muller JP, Elvidge CD. Nighttime imagery as a tool for globalmapping of socioeconomic parameters and greenhouse gas emissions. AMBIO2000;29(3):157–62.

[14] Oda T, Maksyutov S. A very high-resolution global fossil fuel CO2 emissioninventory derived using a point source database and satellite observations ofnighttime lights, 1980–2007. Atmos Chem Phys 2010;10(7):16307–44.

[15] Ghosh T, Elvidge CD, Sutton PC, Baugh KE, Ziskin D, Tuttle BT. Creating a globalgrid of distributed fossil fuel CO2 emissions from nighttime satellite imagery.Energies 2010;3(12):1895–913.

[16] Elvidge CD, Baugh KE, Sutton PC, Bhaduri B, Tuttle BT, Ghosh T, et al. Who’s inthe dark: satellite-based estimates of electrification rates. In: Yang X, editor.Urban remote sensing: monitoring, synthesis and modeling in the urbanenvironment. Chichester, UK: Wiley-Blackwell; 2010. p. 211–23.

[17] Doll CNH, Pachauri S. Estimating rural populations without access toelectricity in developing countries through night-time light satellite imagery.Energy Policy 2010;38(10):5661–70.

[18] NOAA-NGDC. Version 4 DMSP–OLS nighttime lights time series, <http://ngdc.noaa.gov/eog/>; 2010 [accessed 04.09.14].

[19] Chen X, Nordhaus WD. The value of luminosity data as a proxy for economicstatistics, <http://www.nber.org/papers/w16317.pdf>; 2010 [accessed04.09.14].

[20] Elvidge CD, Baugh KE, Kihn EA, Kroehl HW, Davis ER. Mapping city lights withnighttime data from the DMSP operational linescan system. Photogramm EngRemote Sens 1997;63(6):727–34.

[21] State Statistical Bureau, P. R. China. China Energy Statistical Yearbook 2000.Beijing: China Statistical Publishing House; 2001.

[22] IPCC (Intergovernmental Panel on Climate Change). IPCC guidelines fornational greenhouse gas inventories. Hayama, Japan: Institute for GlobalEnvironmental Strategies (IGES); 2006.

[23] Imhoff ML, Lawrence WT, Stutzer DC, Elvidge CD. A technique for usingcomposite DMSP/OLS ‘‘city lights’’ satellite data to map urban area. RemoteSens Environ 1997;61(3):361–70.

[24] Small C, Pozzi FCD, Elvidge CD. Spatial analysis of global urban extent fromDMSP–OLS night lights. Remote Sens Environ 2005;96(3):277–91.

[25] Chunyang H, Peijun S, Jinggang L, Jin C, Yaozhong P, Jing L, et al. Restoringurbanization process in China in the 1990s by using non-radiance-calibratedDMSP/OLS nighttime light imagery and statistical data. Chin Sci Bull2006;51(13):1614–20.

[26] Elvidge CD, Cinzano P, Pettit DR, Arvesen J, Sutton P, Small C, et al. The Nightsatmission concept. Int J Remote Sens 2007;28:2645–70.

[27] United Nations Development Programme. Human development report2011. New York: Palgrave Macmillan; 2011.

[28] Data Sharing Infrastructure of Earth System Science. 1 km grid GDP dataset,<http://www.geodata.cn/Portal/metadata/viewMetadata.jsp?id=100101-25>;2010 [accessed 04.09.14].

[29] Feng ZM, Tang Y, Yang YZ, Zhang D. The relief degree of land surface in Chinaand its correlation with population distribution. Acta Geogr Sin2007;62(10):1073–82.

[30] Kyle W J. The human bioclimate of Hong Kong. In: Kolar R, Kolar M, editors.Contemporary climatology. Proceedings of the COC/IGU meeting, 15–20August, 1994, Brno, Czech Republic, Masaryk University; 1994. p. 345–50.

[31] Thom EC. The discomfort index. Weatherwise 1959;12:57–60.[32] Unger J. Comparisons of urban and rural bioclimatological conditions in the

case of a Central European city. Int J Biometeorol 1999;43:139–44.[33] China Meteorological Administration National Meteorological Information

Center. China surface climate dataset, <http://cdc.cma.gov.cn/home.do>; 2010[accessed 04.09.14].

[34] SPOT-VEGETATION programme. SPOT-VEGETATION programme data, <http://www.spot-vegetation.com/>; 2010 [accessed 04.09.14].

306 H. Lu, G. Liu / Applied Energy 131 (2014) 297–306

[35] State Environmental Protection Administration of China. Environment ProtectionIndustry Criterion of P.R. China. Technical criterion for eco-environmental statusevaluation. Beijing: China Environment Science Press; 2006.

[36] Cliff AD, Ord JK. Spatial autocorrelation. London: Pion Ltd.; 1973.[37] Cliff AD, Ord JK. Spatial processes: models and applications. London: Pion Ltd.;

1981.[38] Anselin L. Exploring spatial data with GeoDa: a workbook. Urbana-

Champaign: Spatial Analysis Laboratory, Department of Geography,University of Illinois; 2005.

[39] Anselin L. Local indicators of spatial association-LISA. Geogr Anal1995;27:93–115.

[40] Aroca P, Bosch M, Hewings GJD. Regional growth and convergence in Chile1960–1998: the role of public and foreign direct investment. Working Paper ofIDEAR, Universidad Católica del Norte, Valparaíso, Chile; 2001.

[41] Noorman KJ. Exploring futures from an energy perspective – a natural capitalaccounting model study into the long-term economic development potentialof the Netherlands. Doctorial Thesis, Groningen, IVEM, Centrum voor Energieen Milieukunde; 1995.

[42] Hall CAS, Cleveland CJ, Kaufmann RK. The ecology of the economic process,energy, and resource quality. New York: John Wiley & Sons Inc.; 1986.

[43] Trust Carbon. The carbon emissions generated in all that weconsume. London: Carbon Trust; 2006.

[44] Jackson T, Papathanasopoulou E, Bradley P, Druckman A. Attributing UKcarbon emissions to functional consumer needs: methodology and pilotresults. RESOLVE Working Paper 01-07, University of Surrey; 2007.

[45] Brownsword RA, Fleming PD, Powell JC, Pearsall N. Sustainable cities –modelling urban energy supply and demand. Appl Energy 2005;82:167–80.

[46] Mori Y, Kikegawa Y, Uchida H. A model for detailed evaluation of fossil-energysaving by utilizing unused but possible energy-sources on a city scale. ApplEnergy 2007;84(9):921–35.

[47] Vaona Andrea. The sclerosis of regional electricity intensities in Italy: anaggregate and sectoral analysis. Appl Energy 2013;104:880–9.

[48] Li YP, Huang GH, Chen X. Planning regional energy system in association withgreenhouse gas mitigation under uncertainty. Appl Energy 2011;88:599–611.