Spatial distribution of building energy use in the United States through satellite imagery of the earth at night

Building and Environment 142 (2018) 252–264

Contents lists available at ScienceDirect

Building and Environment journal homepage: www.elsevier.com/locate/buildenv

Spatial distribution of building energy use in the United States through satellite imagery of the earth at night

T

Derek Fehrer, Moncef Krarti∗ Building Systems Program, CEAE Department CB 428, University of Colorado, Boulder, CO, 80309, USA

A R T I C LE I N FO

A B S T R A C T

Keywords: Building energy use Nighttime lights Satellite images Spatial and temporal energy use maps

Despite the importance of geospatial analysis of energy use in buildings, the data available for such exercises is limited. A potential solution is to use geospatial information, such as that obtained from satellites, to disaggregate building energy use data to a more useful scale. Many researchers have used satellite imagery to estimate the extent of human activities, including building energy use and population distribution. Much of the reported work has been carried out in rapidly developing countries such as India and China where urban development is dynamic and not always easy to measure. In countries with less rapid urbanization, such as the United States, there is still value in using satellite imagery to estimate building energy use for the purposes of identifying energy efficiency opportunities and planning electricity transmission. This study evaluates nighttime light imagery obtained from the VIIRS instrument aboard the SUOMI NPP satellite as a predictor of building energy use intensity within states, counties, and cities in the United States. It is found that nighttime lights can explain upwards of 90% of the variability in energy consumption in the United States, depending on conditions and geospatial scale. The results of this research are used to generate electricity and fuel consumption maps of the United States with a resolution of less than 200 square meters. The methodologies undertaken in this study can be replicated globally to create more opportunities for geospatial energy analysis without the hurdles often associated with disaggregated building energy use data collection.

1. Introduction The latest IPCC report suggests that global annual greenhouse gas (GHG) emissions must decrease 40–70% by 2050 and be entirely neutralized by 2100 to keep global temperatures from increasing more than 2 °C and to avoid the worst consequences of climate change. Globally, buildings are responsible for a third of all energy consumption and greenhouse gas (GHG) emissions [1]. In the United States, buildings are responsible for 40% of energy consumption and GHG emissions [2]. In order to meet the IPCC's goals, energy use and emissions from the building sector must be reduced substantially. Understanding where and how buildings consume energy is important for identifying opportunities for energy use reductions developing efficient electricity distribution networks, where in the US for example, buildings consume 70% of electricity [3]. Despite the importance of geospatial analysis of energy use in buildings, the data available for such exercises is limited. Building energy use data at the state level is obtainable from the Energy Information Administration in the US [4]; however, building energy use data at finer scales is difficult to find because it is held by many different parties and often access is

∗

Corresponding author. E-mail address: [email protected] (M. Krarti).

https://doi.org/10.1016/j.buildenv.2018.06.033 Received 28 December 2017; Received in revised form 17 May 2018; Accepted 13 June 2018 Available online 15 June 2018 0360-1323/ © 2018 Elsevier Ltd. All rights reserved.

restricted due to privacy concerns. Methods for obtaining or estimating building energy use at finer scales utilize bottom-up or top-down approaches. Bottom-up approaches focus on energy consumed in individual buildings and often employ statistical and deterministic energy models that account for physical building characteristics [5]. These approaches require large samples of energy use data and building information in order to calibrate the models. National surveys of building energy use, such as CBECS [6] and RECS [7], are a useful resource for bottom-up approaches to modeling building energy use, but extending the model beyond the sample buildings requires detailed information about the entire building stock, which is not usually available. Utility and energy supply companies have information on customer-specific building energy use derived from either utility bills or meter data; however, this information is not readily available to third parties due to privacy concerns. In fact, several US states have laws that protect customer privacy and limit what information utilities can share with third parties [8,9,10]. Top-down modeling of energy consumption generally consists of establishing relationships between energy consumption at a coarser

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

nighttime lights and socio-economic indicators at multiple scales. Although, some studies found that the relationship is not as strong at smaller scales [10,20] and in areas with specific activities, such as mining, which increasenighttime lights without a commensurate increase in socio-economic indicators like population [25]. In the study summarized in this paper, the suitability of nighttime lights as an indicator of electricity and stationary fuel consumption, including natural gas and heating oil, is evaluated at three geospatial scales and compared against other available socio-economic datasets. The potential for disaggregating energy data with nighttime light satellite imagery is also explored. The general methodology is first outlined. Then, applications of the nighttime lights to predict site and source building energy uses for states, counties, and cities within the contiguous US are described and select results are discussed.

scale with variables that impact energy consumption, such as population, GDP, or satellite imagery. Top-down approaches typically face fewer privacy-related regulatory hurdles because the requisite data are at such a high level that they do not reveal specific customer information. These methods also do not rely on specific building energy consumption for model calibration. Top-down allocation of energy consumption can be achieved with census data, economic data, or property tax assessment records [11,12]. However, these methods are limited in their applicability at small scales and replicability across geographies. Census data may not achieve accurate high-resolution disaggregation as the location of domiciles does not capture commercial and industrial energy use that occurs in city centers or areas with few permanent residents. Economic data is often not available at small scales. Property tax assessment records reside in different databases held by multiple organizations, necessitating coordination between multiple parties especially across larger geographic regions. Unlike the aforementioned methods, satellite imagery of the earth at night, which shows the extent and intensity of human activity, provides a scalable and replicable dataset for top-down geospatial disaggregation of energy consumption. Additionally, this imagery-based data is useful for mapping and quantifying the light pollution associated with anthropogenic light sources and investigating the impacts of this pollution on human health [13,14,15]. Nighttime light imagery is available from both the DMSP and VIIRS satellites, which travel in sun-synchronous orbits and measure the radiance of the earth during both the daytime and nighttime pass. The measurements taken during the nighttime pass include non-anthropogenic sources of light such as lightning, fires, and reflected moonlight. The Earth Observation Group at the NOAA National Geophysical Data Center processes the raw data, removing many sources of non-anthropogenic lights, including cloud-cover and ephemeral lights, and generates images showing mainly the lights from human sources. Annual composite stable-lights imagery from the DMSP satellite is available from 1992 to 2013. Beginning in 2012, imagery from the Visible Infrared Imaging Radiometer Suite Day Night Band (VIIRS DNB) scanning radiometer aboard the Suomi National Polar-Orbiting Partnership (NPP) satellite was made available. The VIIRS imagery offers many benefits over the DMSP imagery. The VIIRS DNB has a constant spatial resolution of 742 m × 742 m [16,17] compared to DMSP-OLS which has a ground footprint of 5 km × 5 km [18]. VIIRS imagery is projected to a 15 arcsecond grid compared to a 30 arcsecond grid for the DMSP-OLS imagery. The upper end of the dynamic range for the DMSP-OLS imagery is 10−8 W cm−2·sr−1 [19], and the sensor often saturated when measuring densely lit urban cores. Conversely, the VIIRS DNB has a larger dynamic range of 3 × 10−9 W cm−2·sr−1 to 0.02 W cm−2·sr−1, although the instrument has been found to actually outperform this range with a low end of 5 × 10−11 W cm−2·sr−1 [16], and does not saturate. An onboard solar diffuser is used to calibrate the VIIRS DNB measurements so that radiance can be reported in units of W·cm−2·sr−1 whereas the DMSP-OLS measurement were taken only as a digital number from 0 to 63 that required post-processing to calibrate to an actual radiance value. The VIIRS DNB instrument collects panchromatic radiometric data in the range of 0.5–0.9 μm. Because the VIIRS-DNB imagery is relatively new, most of the literature using nighttime lights as a measure of socio-economic activity has considered imagery from the DMSP-OLS. A significant portion of the reported research was focused on correcting for the shortcomings of the DMSP-OLS imagery, such as developing methods to overcome the low-resolution [20], post-processing calibration [21,22], and correcting for over-saturated pixels in city centers [23]. While the VIIRS-DNB imagery does not have many of these shortcomings, the images can still benefit from processing to deal with over glow and seasonal variations in radiance [20,24]. Nighttime lights has been used to measure urban extent [20,25], population [22,26], economic output [19,24], and energy use [25,10,27,19]. These studies found a strong relationship between

2. Methodology description 2.1. Overview The images used for this study were monthly VIIRS DNB composite from 2012 through 2016 and an annual composite of 2015 (“VIIRS Cloud Mask - Outlier Removed - Nighttime Lights”) which was processed by NOAA, as described by Baugh et al. and Elvidge et al., to remove outliers and to set background non-lights to zero [28,29]. Nighttime light measurements were compared to building energy use at the state-level and city-level in the contiguous United States and at the county-level in California. Building energy use was evaluated using four different scopes: electricity, stationary fuel, site energy, and source energy. Site energy is the energy consumed at the point of use, e.g. electricity consumed by a building. Source energy is the total upstream energy consumed to provide a unit of site energy, e.g. the coal consumed in a powerplant to generate the electricity used by the building. Source energy was calculated using EPA Portfolio Manager coefficients of 3.14 for electricity and 1.05 for natural gas [30]. Additionally, other potential predictors of building energy use including GDP, population, latitude, land area, elevation, cooling degree days (CDD), and heating degree days (HDD) were evaluated for their potential to predict electricity and stationary fuel use. 2.2. Total night lights (TNL) Total night lights (TNL) is a common approach for quantifying the nighttime lights in a region [31,19], this is also sometimes referred to as “sum of lights” [32,33]. In this paper, the convention “total night lights” or TNL is used. VIIRS DNB images were analyzed using the WGS 84 projection system and the TNL for each region was calculated using an area-weighted sum of all radiance measurements within a polygon defining the region (see equation (1)). K

TNL =

A

∑ Lk ∗ Ak k=1

(1)

R 2

In Eq. (1), Lk is the radiance measured by the scanner in nW/cm -sr, Ak is the area of pixel, and AR is the reference area of a 15 arc-second pixel at the equator (463 m × 463 m). This definition of TNL is consistent with other research [23], and does not require making assumptions about surface conditions (e.g. Lambertian) that would be necessary to calculate photometric properties such as radiant intensity or exitance. Most of the satellite images were not filtered or modified before use. While some researchers have patched and smoothed the data from VIIRS-DNB, and set a minimum threshold of 9 nW/cm2-sr, these processes provided only marginal improvements to the regression analysis and were not undertaken for this analysis [24]. Additionally, setting a threshold of 9 nW/cm2-sr would remove valid anthropogenic light sources from the image, for example the San Mateo Bridge was found to have an average radiance of ∼4 nW/cm2-sr by Ref. [16]. 253

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

Fig. 1. Regression analysis of annual electricity consumption and total night lights in the US by state.

performed to evaluate how additional predictors, besides TNL, could increase the accuracy of the model. An exhaustive best-subset selection process was conducted using the regsubsets function in R to identify the best predictors for estimating building energy use.

To obtain annual average TNL values from monthly composite images, the TNL for all regions in the monthly images were calculated and then averaged to obtain an annual average TNL for each region. As Zhao et al. found [24], higher latitudes do not have usable measurements in the summer months due to stray light interfering with the measurement. Review of the number of pixels in each month that had zero cloud-free observations revealed that this effect is limited to April through August. This must be handled prior to the regression analysis as pixels with zero cloud-free observations are noted as having a radiance value of zero in the composite images. Rather than apply corrective algorithms as Zhao et al. have done [24], these months containing many pixels with no usable observations were simply excluded from the analysis for all states. In the analysis of California, the months with many missing pixels were limited to May through July, so for the county-level analysis only those months were removed.

y = β1 x 1 + β2 x 2 + β3 x 3 + …+βp x p

(3) 2

The coefficient of determination, R value, was used as the measure for the performance of the model in fitting to the data as indicated by Eq. (4):

R2 = 1 −

ˆ ∑ (y − y )

∑i (yi − yi ) i

i

2

2

(4)

3. Discussion of main results 3.1. State-level analysis

2.3. Regression analysis Retail sales of electricity to the commercial, industrial, and residential sectors were obtained from the EIA for each state for the years 2012–2016. Retail sales of natural gas, distillate fuel oil, residual fuel oil, and heating kerosene were obtained from EIA for each state for the same period, converted to GWh, and aggregated as fuel consumption. Alaska was excluded from the analysis because it was found to be an outlier, likely due to its large size and low population density, and Hawaii was excluded because of insufficient data. Washington DC was included and treated as a state for a total of 49 regions in the analysis. The average land area of the states included in the analysis is 156,187 km2 (60,304 mi2).

Ordinary least squares regression analysis was used to examine the relationship between TNL and stationary building energy use. Linear regression makes it easier to directly compare the relationship between nighttime lights and building energy use across different regions. Previous studies have found a linear model to describe the relationship between TNL and building energy use well [22] and examination of the residuals from regression analysis for this study found a linear relationship to be adequate. Additionally, the p-value of the constant value coefficient for the regression analysis was always found to be greater than 5%. This indicated that the constant coefficient was not statistically significant, and because the expectation was that there is no building energy use where there are no nighttime lights, a single variable regression analysis was used to relate building energy use and TNL, as shown in Eq. (2), where y equals energy consumption and x equals TNL. With a single coefficient, the regression analysis can be conceptually simplified to the slope of the line (β1) being equal to the amount of energy consumed per unit of nighttime light.

y = β1 ∗x

3.1.1. Multi-year linear regression Fig. 1 shows annual electricity consumption for each state plotted against TNL with a best-fit linear model. The slope of the best-fit line ranges from 0.0785 to 0.0877 for individual years and is 0.083 for all years together. The R2 value ranges from 0.879 to 0.928 for individual years and is 0.904 for all years together. This indicates the model is a good fit for the data. The results from the linear regression analysis, including regressions against fuel use, site energy, and source energy, are summarized in

(2)

In addition, multivariable regressions as shown in Eq. (3) were 254

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

the reason that less energy is used in 2013–2015 per unit of nighttime light has more to do with changes in the amount of nighttime light than the amount of building energy use.

Table 1 Results of linear regression with a single coefficient, the slope (β1), using US State data. Energy Scope

Year

Slope (β1)

(β1) p-value

R2

Electricity Electricity Electricity Electricity Electricity

2012 2013 2014 2015 2016

0.087 0.081 0.078 0.082 0.088

< 0.001 < 0.001 < 0.001 < 0.001 < 0.001

0.879 0.901 0.903 0.928 0.928

Fuel Fuel Fuel Fuel Fuel

2012 2013 2014 2015 2016

0.118 0.117 0.116 0.114 0.122

< 0.001 < 0.001 < 0.001 < 0.001 < 0.001

0.708 0.735 0.760 0.744 0.737

2012 2013 2014 2015 2016

0.206 0.198 0.194 0.196 0.210

< 0.001 < 0.001 < 0.001 < 0.001 < 0.001

0.835 0.858 0.878 0.878 0.876

2012 2013 2014 2015 2016

0.399 0.378 0.368 0.376 0.404

< 0.001 < 0.001 < 0.001 < 0.001 < 0.001

0.882 0.905 0.918 0.929 0.929

Site Site Site Site Site

energy energy energy energy energy

Source Source Source Source Source

energy energy energy energy energy

3.1.2. Comparison to 2015 “scrubbed” imagery Regression analysis was also performed using “VIIRS Cloud Mask Outlier Removed - Nighttime Lights” data for 2015, which has temporal lights such as fires removed and background non-lights set to zero. This is referred to as “scrubbed” data in this paper. The process of scrubbing the TNL data removed 17% of the nighttime lights found in 2015 using un-scrubbed monthly data. This results in the slope of the regression lines being steeper for the scrubbed data, as summarized in Table 2, compared to the regressions using un-scrubbed data shown in Table 1. Except for site building energy use, the R2 value for the 2015 scrubbed data (Table 2) is surprisingly less than the R2 value for the unscrubbed data (Table 1). This result could be because scrubbing removes some lights associated with activity in the built environment. It could also reflect the fact that northern states have fewer usable observations in the summer months compared to southern states, so in the scrubbed data the annual average for southern states is based on all months of the year while the annual average for northern states essentially excludes April through August. The scrubbed data introduce bias and could change the relationship that each state has between building energy use and TNL. The method that Zhao [24] developed to generate annual nighttime light images from monthly images may be preferable to the “VIIRS Cloud Mask - Outlier Removed - Nighttime Lights” data product. Alternatively, an annual “VIIRS Cloud Mask Outlier Removed - Nighttime Lights” that has undergone NOAA's stray lights removal procedure might reduce the bias caused by fewer cloud free observations at northern latitudes, but this product is not currently available.

Table 1. The slope of the regression line for fuel use is steeper than electricity consumption, because more fuel than electricity is consumed in the commercial, industrial, and residential sectors in the US. The lower R2 value of the fuel regression indicates that TNL does not fit fuel use as well as it fits electricity use. This makes sense as nighttime lights are expected to be generated from electrically-powered sources more commonly than fuel-powered sources, although some gas flares are captured in the data. Additionally, fuel use for heating is highly dependent on the climate zone. TNL versus source energy results in the best fit. Buildings in warm climates are likely to use more electricity because of space cooling requirements. The opposite is true for natural gas, where buildings in cold climate are likely to use more fuel because of space heating requirements. Performing a regression analysis on TNL versus only electricity or fuel does not capture this shift in building energy use types between warm and cold climates. Site energy captures this trade-off, but does not account for the fact that cooling is most frequently achieved with heat pumps that are more effective at providing cooling than the furnaces and boilers that most often provide heating. The 3× multiplier used to convert site electricity use to source electricity use could help account for this, resulting in a better fit for source energy than site energy. The better fit between TNL and source energy compared to other metrics of building energy use is supported by data obtained from CBECS. An analysis of energy use intensities of 18 building types for 5 climate zones found that the coefficient of variance (COV) for source energy intensity was lower than the electricity use intensity COV for 14 of the 18 building types and lower than the site energy intensity COV for 12 of the 18 building types. The slope of the regression line for each year is shown in Fig. 2. The COV for the slope over time ranges from 3% for fuel to 6% for electricity. Aggregated electricity and stationary fuel consumption is shown in Fig. 3, along with average state-level HDD and CDD, for available years, and the aggregate TNL. This helps shed light on why the relationship between building energy use and TNL changes over time. While the average HDD and CDD by state do not account for area weighting, they do indicate general climatic trends and suggests the reason for the increase in fuel consumption during 2013 and 2014 is due to increased space heating demand. The shape of the curves in Fig. 2 are roughly the inverse of the TNL curve in Fig. 3, suggesting that

3.2. County-level analysis California was chosen for county-level analysis because there is publicly available information on electricity and natural gas consumption by county for the years 2012 through 2016, provided by the California Energy Commission. It was assumed that no other fuels were consumed in buildings in any significant amount. All 58 California counties were considered in the analysis. The average land area of the counties is 6.9 × 106 m2 (2686 mi2). During regression analysis, it was found that the fuel data for California has a few significant outliers. Kern County was the largest outlier, consuming roughly 70,000 GWh of natural gas per year while having an annual average total night lights value of only approximately 150,000 nW/cm2sr. This anomaly may be because Kern County is a large producer of oil and gas. Non-residential consumption of natural gas is 30 times higher than residential use in Kern County, compared to the state average of 2. This is supported by Kern County's GHG inventory, which reports that in 2005 natural gas and waste gas emissions from the refining industry were roughly 18 times higher than residential emissions from natural gas use [34]. The second largest outlier county was Contra Costa County which consumed roughly 30,000 GWh of natural gas per year and had an annual average total night lights value of approximately 75,000 nW/ cm2sr. Contra Costa County used roughly six times as much natural gas for non-residential purposes as it did residential, potentially due to the many oil refineries in the county. Removing Contra Costa and Kern counties from the analysis substantially improved the R2 of the linear regression, and so these two counties were not considered in further regression analysis. These outliers suggest that nighttime lights are not well suited to account for natural gas consumed in fossil fuel extracting and refining. 3.2.1. Multi-year regression analysis Fig. 4 shows annual electricity consumption for each county plotted against the total night lights for the county with a best-fit linear model. 255

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

Fig. 2. Slope of the regression analysis for the US states for different building energy use scopes.

All regression models show a good fit with R2 greater than 0.96. The slope of the best-fit line ranges from 0.0973 to 0.107 for individual years and is 0.966 for all years together. This is 12% higher than the slope of the line for the United States as a whole, which was 0.083. This finding means California consumes 12% more electricity per unit of TNL than the average US state. This is surprising because, on a per capita basis, California consumes much less energy than other states. However, California also has fewer nighttime lights per person than the average state. According to the Building Performance Database from Lawrence Berkeley Laboratory [35], the electricity use intensity for commercial buildings in California (79 kBtu/ft2-year) is similar to the national average (80 kBtu/ft2-year). This would indicate that California uses more electricity per unit of nighttime light not because it is generally more electricity intensive than other states, but because it has fewer nighttime lights. The results from the linear regression analysis, including regressions against fuel use, site energy, and source energy, are summarized in Table 3. The slope of the best-fit line ranges from 0.102 to 0.123 for individual years and is 0.11 for all years together. This is 6% lower than the slope of the line for the United States as a whole, which was 0.117. This suggests that in California, buildings consume slightly less fuel per unit of TNL than the average US state. The R2 value for linear regression

Table 2 Results of linear regression with a single coefficient, the slope (β1), using US State data for 2015 that has been scrubbed by NOAA. Energy Scope

Year

Slope (β1)

(β1) p-value

R2

Electricity Fuel Site energy Source energy

2015 2015 2015 2015

0.087 0.139 0.226 0.527

< 0.001 < 0.001 < 0.001 < 0.001

0.846 0.741 0.897 0.821

modeling fuel consumption ranges from 0.92 to 0.936 for individual years and is 0.923 for all years together. This is a better fit than the model trained based on data from US states, which had an R2 value of 0.741. This may be because states can have very different heating load profiles due to different climates, whereas California's climate is relatively uniform. The slopes of the regression lines over time are plotted in Fig. 5. The changes are relatively small, with the COV ranging only from 4% for electricity to 8% for natural gas. Annual variation of both electricity and fuel consumption for California for the period of 2012 through 2016 is shown in Fig. 6. Electricity use is relatively constant while fuel use is more variable with a peak in 2013. There appears to be some

Fig. 3. (A) Annual electricity and fuel consumption as site energy in the United States for 2012–2016, (b) mean US state heating degree days and cooling degree days for 2012–2015, (c) Total night lights in US for 2012–2016. 256

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

Fig. 4. Regression analysis of annual electricity consumption and total night lights in California by county.

3.2.2. Comparison to 2015 “scrubbed” imagery Regression analysis for building energy use in California counties in 2015 was also performed using “VIIRS Cloud Mask - Outlier Removed Nighttime Lights” data, which has temporal lights such as fires removed and background non-lights set to zero. The “scrubbed” TNL values with outliers removed and background lights set to zero were found to be 9% lower than the data used for the multi-year analysis. This results in the slope of the regression lines being steeper for the scrubbed data compared to the un-scrubbed data. The R2 values for the 2015 scrubbed data (Table 4) are marginally higher than the R2 values for corresponding un-scrubbed data (Table 3). This is the opposite of the US States data, where the un-scrubbed data had a better fit than the scrubbed data. There were a number of wildfires in 2015 in California, which could explain why removing temporal lights improved the goodness of fit between building energy use and total night lights. A comparison of the difference between the TNL calculated based on scrubbed vs. un-scrubbed data found that of the 10 counties with the largest difference, 7 had wildfires in 2015, as shown in Table 5.

Table 3 Results of linear regression with a single coefficient, the slope (β1), for California counties. Energy Scope

Year

Slope (β1)

(β1) p-value

R2

Electricity Electricity Electricity Electricity Electricity

2012 2013 2014 2015 2016

0.1072 0.0976 0.0979 0.0973 0.0986

< 0.001 < 0.001 < 0.001 < 0.001 < 0.001

0.967 0.967 0.964 0.967 0.970

Fuel Fuel Fuel Fuel Fuel

2012 2013 2014 2015 2016

0.123 0.116 0.103 0.102 0.108

< 0.001 < 0.001 < 0.001 < 0.001 < 0.001

0.927 0.928 0.920 0.929 0.938

2012 2013 2014 2015 2016

0.230 0.214 0.201 0.199 0.207

< 0.001 < 0.001 < 0.001 < 0.001 < 0.001

0.953 0.953 0.951 0.956 0.961

2012 2013 2014 2015 2016

0.466 0.429 0.416 0.413 0.423

< 0.001 < 0.001 < 0.001 < 0.001 < 0.001

0.962 0.962 0.960 0.964 0.968

Site Site Site Site Site

energy energy energy energy energy

Source Source Source Source Source

energy energy energy energy energy

3.3. City-level analysis In the US, the most comprehensive database of building energy use at the city-level is provided by the US Department of Energy's Cities Leading through Energy Analysis and Planning (Cities-LEAP) and State and Local Energy Data (SLED) projects [3]. As part of this program, NREL created energy profiles for over 23,000 cities in the United States based on year 2013 data. The energy use was derived for each city through analytical, spatial refinement processes for the residential, commercial, and industrial sectors whereby state level consumption data was allocated to the city level. It should be noted that the cities in the SLED database are what the US Census Bureau refers to as places. For the sake of clarity, these US Census Bureau places are primarily referred to as cities throughout this text. The energy consumption information from SLED is available through a web interface that displays the data for one city at a time; no option exists to download data for multiple cities. Obtaining city data manually one-by-one was infeasible, so an API query was built using R

correlation between the heating degree days shown in Fig. 6 and the fuel use shown in Fig. 6, but clearly heating degree days are not the only determinant of fuel use. The curves in Fig. 5 are roughly the inverse of the annual TNL plotted in Fig. 6, which is similar to the annual TNL plotted for US States in Fig. 3. This suggests that, as with the state-level analysis, the change in slope relating TNL to building energy use is due more to changes in TNL than changes in building energy use. It follows that there are limits to the usefulness of TNL for tracking building energy use changes over time unless temporal changes in the satellite measurements are accounted for.

257

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

Fig. 5. Slope of the regression analysis for the US states for different building energy use scopes.

to download the consumption data for each city programmatically. A list of US Census Bureau Places with a “C” or “U” class code, generally representing incorporated places and census designated places, respectively, was obtained from US Census Bureau TIGER shapefiles. Of the 28,685 places from the US Census Bureau shapefile, electricity and natural gas data from the SLED database were obtained for 19,281. The remaining 9404 places presumably were not available or had different names in the SLED database.

Table 4 Results of linear regression with a single coefficient, the slope (β1), using CA county data for 2015 that has been scrubbed by NOAA.

3.3.1. Single-year regression Year 2013 TNL values for each city were determined by calculating and then averaging the monthly TNL for September through March. Regression analysis was performed using a single coefficient for all energy scopes, with select results shown in Fig. 7 and all results summarized in Table 6. The regression against electricity provides a good fit with an R2 value of 0.832. The R2 value for the regression against fuel is lower at 0.627. The largest cities in terms of total night lights are, in order, Chicago, Houston, New York, Los Angeles, Dallas, and Phoenix. Of these cities, New York is a clear outlier when it comes to electricity use, as it consumes nearly 150% more electricity than any other city even though it is not the most brightly lit. The electricity consumption for New York City from the SLED database is similar to what is reported

Energy Scope

Year

Slope (β1)

(β1) p-value

R2

Electricity Fuel Site energy Source energy

2015 2015 2015 2015

0.100 0.108 0.208 0.428

< 0.001 < 0.001 < 0.001 < 0.001

0.976 0.943 0.967 0.974

by the NYISO for electricity consumption, so the electricity consumption from SLED appears to be valid [36]. The R2 values for the city-level analyses are lower than the values for the state-level analyses. This is similar to what other researchers have reported [10,20] and may be because smaller geospatial regions have greater variability that is averaged out at a larger scale, such as at the state-level. The coefficient (β1) for the city-level analysis is 27% higher for electricity and 22% higher for fuel consumption compared to the regression analysis performed at the state-level. The steeper slope of the line for the city-level regression means that in the US, cities have more building energy use per unit of TNL than states. It is important to note

Fig. 6. (A) Annual electricity and fuel consumption as site energy in California for 2012–2016, (b) California heating degree days and cooling degree days for 2012–2015, (c) Total night lights in California for 2012–2016. 258

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

Table 5 Counties with a large difference between 2015 TNL based on un-scrubbed and scrubbed datasets and which of these counties had wildfires. County

San Bernardino Ventura Riverside Fresno Inyo Imperial Tulare Siskiyou Lassen Trinity

TNL Unscrubbed (nW/ cm2-sr)

TNL Scrubbed (nW/cm2-sr)

Difference (nW/cm2-sr)

Wildfire

240,271 80,874 212,052 82,269 14,623 36,467 43,775 8757 9524 4957

210,087 55,231 198,518 69,553 1910 27,324 35,079 2152 4555 376

30,184 25,643 13,534 12,716 12,714 9143 8696 6606 4968 4581

X X X X X

Table 6 Results of linear regression with a single coefficient, the slope (β1), for US cities. Energy Scope

Year

Slope (β1)

(β1) p-value

R2

Electricity Fuel Site energy Source energy

2013 2013 2013 2013

0.103 0.141 0.244 0.472

< 0.001 < 0.001 < 0.001 < 0.001

0.833 0.611 0.732 0.794

Table 7 Single coefficient regression characteristics for city-level analysis compared to state-level analysis. Energy Scope

Year

X X Electricity Fuel Site energy Source energy

that while the area of the contiguous United States is 8 million km2, the area of the incorporated places and census designated places (cities) is only 436,000 km2, or roughly 5.5% of the country's area. Yet these regions held over 196 M people in 2013 or 62% of the US total population and are therefore much denser with human activity. The higher building energy use per unit of nighttime light could be due to relatively fewer non-anthropogenic light sources in dense regions compared to sparsely populated regions. For example, the state of California has vast unpopulated areas where relatively small amounts of background light present over a large area could make a significant impact on the total night lights for the states, whereas the cities in California are not impacted as much by these small amount of background lights.

2013 2013 2013 2013

City

State

Slope (β1)

R2

Slope (β1)

R2

0.103 0.141 0.244 0.472

0.833 0.611 0.732 0.794

0.081 0.117 0.198 0.378

0.902 0.739 0.861 0.906

8 predictors in the top row. The gradient of the colored box, from grey to black, indicates the R2 value of the linear model. The best single predictor for both electricity and fuel use is population. In the case of electricity consumption, the performance of the model improves significantly by adding TNL divided by latitude to the model as a regressor. This potentially reflects higher cooling loads and electricity consumption at lower latitudes compared to higher latitudes. Similarly, adding TNL multiplied by the latitude improves the fuel consumption model as it reflects increased heating loads at higher latitudes. Including more than two regressors does not significantly improve the accuracy of the model. These results show how adding an indicator of regional climate can improve regression analysis using TNL to estimate building energy use.

3.3.2. Other predictors of building energy use Additional factors that are likely to impact building energy use were included in a best subset collection process. The factors included are listed here:

3.3.3. Multi-state analysis A series of 49, single-coefficient linear regressions were performed using city-level data, one for each of the contiguous United States and the District of Columbia. When the city-level regression is performed separately for each state as opposed to all the cities together, the goodness of fit improves significantly. While the R2 value for all cities together was 0.833 for electricity and 0.611 for fuel, the average R2 value of the city-level regressions performed individually for each state were 0.915 for electricity and 0.783 for natural gas. This improvement is clear in the histogram displayed in Fig. 9, which shows that the

• Population: population of each city from the 2013 census • Land area: land area (excluding water) in m for each city • Latitude: latitude of each city's centroid 2

The regressors listed above, along with TNL, were considered independently and in combination for their ability to predict electricity and stationary fuel use at the city-level. The results are shown in Fig. 8, in which each row represents a specific subset of regressors, indicated by the colored boxes, ranging from 2 predictors in the bottom row up to

Fig. 7. Results of linear regression with a single coefficient for US cities. 259

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

Fig. 8. Results of best subset selection of prediction parameters for city electricity and fuel use for 2013.

Fig. 10. Distribution of the coefficient relating building energy use to TNL based on 2013, city-level, single coefficient regressions performed independently for each state (bars) compared to the regressions performed for all cities in the US together (vertical dashed lines).

Fig. 9. Distribution of the R2 value for city-level, single coefficient regressions performed independently for each state (bars) compared to the R2 value of regressions performed for all cities in the US together (vertical dashed lines).

majority of regressions performed at the city-level independently for each state (the bars) have higher R2 values than city-level regressions performed for all cities together (the dashed lines). Fig. 10 shows the distribution of the single-coefficients defining the relationships between electricity use and TNL and between stationary fuel use and TNL for the city-level analysis performed independently for each state (the bars) versus the analysis carried out for all states together (the dashed lines). The states with the highest slope for electricity, indicating more electricity use per unit of TNL, are New York, Washington, Oregon and West Virginia. The states with the lowest slope for electricity, indicating the least amount of consumption per unit of TNL, are North Dakota, South Dakota, Montana, and Wisconsin. The slope of the relationship for the highest state (New York) is nearly five times greater than the slope of the line for the lowest state (North Dakota). The states with the highest slope for fuel use are New York, Louisiana, California, and Indiana. The states with the lowest slope for fuel use are Florida, Arizona, North Carolina, and North Dakota. The slope of the state with the highest value (New York) is 14 times greater than the state with the lowest value (Florida). These findings indicate that there is high variability in the relationship between nighttime lights and building energy use, and that much of this variability can be managed by examining states separately. Because these regressions are based only on cities, the nighttime lights are more likely attributed to buildings, as opposed to other human activities such as mining or background lights in sparsely populated areas.

3.3.4. City-level analysis compared to state-level analysis While city-level energy consumption data is available for 2013, this is not the case for most years. Therefore, it is valuable to examine whether state-level energy consumption can be accurately allocated to cities based on the TNL. One approach is to determine the relationship between state-level building energy use and TNL and then allocate building energy use within the state based on the distribution of TNL. If performed separately for each state, this accounts for the widely varying relationship between building energy use and TNL in different states, as shown in Fig. 10. The accuracy of this approach was evaluated by comparing the relationship calculated by dividing state building energy use by state TNL against the relationship found by performing regression analysis using the city-level energy and TNL as data points in each state, as shown in Fig. 11. The analysis suggests that calculating the relationship between building energy use and TNL at the state-level results in a lower slope than defining the relationship using regression analysis performed for the cities within the states. This is likely due to nighttime lights at the state-level that are not associated with energy consumption at the same rate as lights in cities, such as background lights, fires, or low-energy intensity human activities. While the “VIIRS Cloud Mask - Outlier Removed - Nighttime Lights” data product has these lights removed, that product is only available for 2015. An alternative approach to removing nighttime lights not associated with energy consumption is to “mask” out low radiance measurements 260

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

coefficients are, on average, 5% higher for the city-derived slopes than the state-derived slopes. The electricity consumption coefficients are tightly grouped around the y=x line, which indicates good agreement between the state-level and city-level slope determinations. The fuel consumption coefficients are, on average, 3% lower for the city-derived slopes than the state-derived slopes. The slopes of the fuel lines are not grouped as tightly around the y=x diagonal line, which is expected given the variability in fuel use. This analysis suggests that after appropriate masking to remove background lights, allocating state-level energy consumption to nighttime lights in the state is a reasonable approach for estimating energy consumption of cities within the state based on each city's TNL. 3.4. Application: energy density map of the United States This study showed that TNL is a reasonable metric to use for disaggregating building energy use at the state-level to the city-level. Further study is needed to determine if this relationship holds true at smaller spatial scales, i.e. allocating city-level energy use to areas inside the city. If this relationship does hold true, then building energy use can be disaggregated based on geospatial distribution of nighttime lights, as shown in Fig. 13 and Fig. 14. The maps show electricity and fuel consumption intensity in MWh per km2 per annum for 2013. Fig. 15 shows electricity and fuel use together with electricity as red and fuel use as blue. The energy consumption density in the maps is derived from citylevel energy use as estimated by NREL in the SLED database. If a given pixel falls within a city with known building energy use, then the city's energy use is assigned proportionately to the pixel as shown in Eq. (5):

Fig. 11. Energy vs. TNL relationship from state-level derivation plotted against relationship from linear regression of relationship for cities within the state.

Ej Ek = Lk ∗ TNLj ∗AR Ak

(5)

Where,

• E is the building energy use of a pixel inside city j, • A is the area of the pixel, • L is the radiance value of the pixel, • E is the energy use of the city, • TNL is the total night lights of the city, and • A is the area of the reference pixel (15 arcseconds x 15 arcseconds k

k

k j

j

R

at the equator).

For areas where energy consumption was not obtained from the SLED database, the energy consumption density is estimated using a state-specific factor based on the relationship found by performing single coefficient linear regression on cities within the state. Building energy use for each pixel is estimated using Eq. (6):

Ek L = β1, s k Ak AR

(6)

Where,

• E is the building energy use of a pixel inside state s, • A is the area of the pixel, • L is the radiance value of the pixel, • β is the slope of the regression analysis for the city energy use against city TNL within state, s, and • A is the area of the reference pixel (15 arcseconds x 15 arcseconds

Fig. 12. Energy vs. TNL relationship from state-level derivation plotted against relationship from linear regression of relationship for cities within the state, after masking out pixels with radiance less than −0.5 nW/cm2-sr.

k

k

k

1,s

as these are more likely to be sources of light not associated with building energy use. Visual inspection of the VIIRS DNB imagery determined that a threshold of 0.5 nW/cm2/sr masked out much of the land area of the United States without masking out cities. This reduced the TNL for the contiguous United States by 13%. Conducting both the state-level calculations and city-level regressions with this masked data determined that the coefficients were much more aligned after masking, as shown in Fig. 12. The Earth Observation Group at NOAA takes a more thorough approach to removing background lights that may improve results [29]. After masking, the electricity consumption

R

at the equator).

These maps make it easy to visualize how and where energy is used throughout the United States. Additionally, because the basis for the maps are pixels of less than 0.25 km2, they can provide insight into building energy use at a high resolution. For example, Fig. 16 shows 2013 electricity and fuel use in Las Vegas, NV, allocated based on TNL. The map shows that the city uses slightly more electricity than fuel in 261

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

Fig. 13. Geospatial disaggregation of electricity use, translated into building energy use density for the contiguous United States.

Fig. 14. Geospatial disaggregation of fuel use, translated into building energy use density for the contiguous United States.

Fig. 15. Geospatial disaggregation of electricity and fuel use, translated into building energy use density for the contiguous United States. Red is electricity and blue is fuel. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

are good predictors of electricity and fuel consumption at various geospatial scales within the United States. Single-coefficient linear regression was found suitable to define the relationship between building energy use and nighttime lights. Generally, TNL was more effective as a predictor of energy consumption at larger geospatial scales than smaller geospatial scales. TNL was also found to be more effective at predicting electricity consumption than stationary fuel consumption. Performing regression analysis separately for cities in different states improved the fit of the prediction models, suggesting that there are different energy

buildings, and that the dense center is where most of the energy is used. It is worth noting that by considering city-specific energy allocation instead of state or country-wide allocation methods, the imagery indicates that Las Vegas uses less energy per unit of TNL than other cities. This is likely because Las Vegas is so brightly lit.

4. Summary and conclusions The study summarized in the paper has shown that nighttime lights 262

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

Fig. 16. 2013 electricity and fuel consumption in Las Vegas, allocated by TNL.

use patterns for buildings in different states. Future analysis should take this fact into account when using TNL as a predictor of energy consumption. Through multi-variable regression analysis, it was determined that TNL becomes a more powerful predictor when paired with other metrics indicative of climate, such as latitude and elevation. Future work should explore how adding climate data to the regression analysis can further improve the scalability and applicability of the results, potentially allowing for visual representations of energy consumption within the US under climate scenarios with increased temperatures. It was determined that state-wide building energy use can be allocated to state-wide TNL and achieve roughly similar results to regression analysis performed at the city-level within the states. This can simplify the process of disaggregating state-level energy consumption in the future. As an application, an energy density map of the United States for the year 2013 was created by allocating building energy use based on the distribution of nighttime lights. This map could be combined with geospatial building information, such as square footage, to create energy maps of building energy use intensities. However, additional work is needed to determine the accuracy of the allocation method at finer geospatial scales. Moreover, various applications of VIIRS DNB imagery for drawing insights on energy consumption can be improved by correcting for fluctuations in the satellite measurements due to seasonal variations in ground conditions, blooming of light out from its point of origin, and variations in atmospheric conditions. Additionally, ground truthing the data at fine spatial scales and combining the imagery with other geospatial information such as land-cover and property assessment records is expected to improve accuracy. However, even without these corrective measures, TNL is a good predictor of energy consumption and can be applied at scales that are not feasible with data from other sources.

1017/CBO9780511793677.016. [2] EIA, How much energy is consumed in U.S. residential and commercial buildings? FAQ - U.S. Energy Information Administration (EIA), (n.d.). https://www.eia.gov/ tools/faqs/faq.php?id=86&t=1 (Accessed 25 October, 2017). [3] NREL, NREL: Technology Deployment - cities-LEAP energy profile Tool includes energy data on more than 23,400 U.S. Cities, (n.d.). https://www.nrel.gov/tech_ deployment/ss_leap.html (Accessed 26 November, 2017). [4] EIA, United States - SEDS, (n.d.). https://www.eia.gov/state/seds/seds-datacomplete.php (Accessed 12 December, 2017). [5] W. Li, Y. Zhou, K. Cetin, J. Eom, Y. Wang, G. Chen, X. Zhang, Modeling urban building energy use: a review of modeling approaches and procedures, Inside Energy 141 (2017) 2445–2457, http://dx.doi.org/10.1016/j.energy.2017.11.071. [6] Energy Information Administration, Commercial buildings energy consumption survey (CBECS), (n.d.). https://www.eia.gov/consumption/commercial/(Accessed 30 November, 2017). [7] Energy Information Administration, Residential energy consumption survey (RECS), (n.d.). https://www.eia.gov/consumption/residential/(Accessed 30 November, 2017). [8] California Public Unitility Commity, Decision adopting rules to provide access to energy usage and usage-related data while protecting privacy of personal data, California, (2014) http://docs.cpuc.ca.gov/PublishedDocs/Published/G000/ M090/K845/90845985.PDF. [9] Colorado Department of Regulatory Agencies, Department of Regulatory Agencies Public Utilities Commission Rules Regulating Electric Utilities 4 CCR 723–3, (2007). [10] A.C. Townsend, D. a. Bruce, The use of night-time lights satellite imagery as a measure of Australia's regional electricity consumption and population distribution, Int. J. Rem. Sens. 31 (2010) 4459–4480, http://dx.doi.org/10.1080/ 01431160903261005. [11] A. Fichera, G. Inturri, P. La Greca, V. Palermo, A model for mapping the energy consumption of buildings, transport and outdoor lighting of neighbourhoods, Cities 55 (2016) 49–60, http://dx.doi.org/10.1016/j.cities.2016.03.011. [12] S.C. Taylor, S.K. Firth, C. Wang, D. Allinson, M. Quddus, P. Smith, Spatial mapping of building energy demand in Great Britain, GCB Bioenergy 6 (2014) 123–135, http://dx.doi.org/10.1111/gcbb.12165. [13] F. Falchi, P. Cinzano, D. Duriscoe, C.C.M. Kyba, C.D. Elvidge, K. Baugh, B. Portnov, N.A. Rybnikova, R. Furgoni, Supplement to: the new world atlas of artificial night sky brightness, GFZ Data Services., Sci. Adv (2016) 1–26. [14] F. Falchi, P. Cinzano, C.D. Elvidge, D.M. Keith, A. Haim, Limiting the impact of light pollution on human health, environment and stellar visibility, J. Environ. Manag. 92 (2011) 2714–2722, http://dx.doi.org/10.1016/j.jenvman.2011.06.029. [15] K. Bobkowska, M. Przyborski, J. Szulwic, A method of selecting light sources from night satellite scenes, 15th Int. Multidiscip. Sci. GeoConference SGEM 2015, 2015, pp. 11–18, , http://dx.doi.org/10.5593/SGEM2015/B52/S20.002. [16] C. Cao, Y. Bai, Quantitative analysis of VIIRS DNB nightlight point source for light power estimation and stability monitoring, Rem. Sens. 6 (2014) 11915–11935, http://dx.doi.org/10.3390/rs61211915. [17] C.F. Schueler, T.F. Lee, S.D. Miller, VIIRS constant spatial-resolution advantages, Int. J. Rem. Sens. 34 (2013) 5761–5777, http://dx.doi.org/10.1080/01431161. 2013.796102. [18] C.D. Elvidge, K. Baugh, M. Zhizhin, F.C. Hsu, Why VIIRS data are superior to DMSP for mapping nighttime lights, Proc. Asia-Pacific Adv. Netw (2013) 62–69, http://dx. doi.org/10.7125/APAN.35.7.

References [1] D. Ürge-Vorsatz, N. Eyre, P. Graham, D. Harvey, E. Hertwich, Y. Jiang, C. Kornevall, M. Majumdar, J.E. McMahon, S. Mirasgedis, S. Murakami, A. Novikova, K. Janda, O. Masera, M. McNeil, K. Petrichenko, S. Tirado Herrero, Energy end-use: buildings, Glob. Energy Assess. Towar. a Sustain. Futur (2012) 649–760, http://dx.doi.org/10.

263

Building and Environment 142 (2018) 252–264

D. Fehrer, M. Krarti

[27] L. Coscieme, F.M. Pulselli, S. Bastianoni, C.D. Elvidge, S. Anderson, P.C. Sutton, A Thermodynamic geography: night-time satellite imagery as a proxy measure of emergy, Ambio 43 (2014) 969–979, http://dx.doi.org/10.1007/s13280-0130468-5. [28] K. Baugh, F.-C. Hsu, C.D. Elvidge, M. Zhizhin, Nighttime lights compositing using the VIIRS day-night band: preliminary results, Proc. Asia-Pacific Adv. Netw 35 (2013) 70, http://dx.doi.org/10.7125/APAN.35.8. [29] C.D. Elvidge, K. Baugh, M. Zhizhin, F.C. Hsu, VIIRS night-time lights, Int. J. Rem. Sens. 38 (2017) 5860–5879, http://dx.doi.org/10.1080/01431161.2017.1342050. [30] Energy Star, Source Energy (2013) 1–17. [31] X. Jing, X. Shao, C. Cao, X. Fu, L. Yan, Comparison between the Suomi-NPP daynight band and DMSP-OLS for correlating socio-economic variables at the provincial level in China, Rem. Sens. 8 (2016) 1–24, http://dx.doi.org/10.3390/ rs8010017. [32] T. Ghosh, C.D. Elvidge, P.C. Sutton, K.E. Baugh, D. Ziskin, B.T. Tuttle, Creating a global grid of distributed fossil fuel CO2 emissions from nighttime satellite imagery, Energies 3 (2010) 1895–1913, http://dx.doi.org/10.3390/en3121895. [33] C.C.M. et a. Kyba, Artificially lit surface of Earth at night increasing in radiance and extent, Sci. Adv (2017) 1–9, http://dx.doi.org/10.1126/sciadv.1701528. [34] S. Joaquin, V. Air, P. Control, K. County, C. Development, Kern County Communitywide Greenhouse Gas Emission Inventory, (2012). [35] Lawrence Berkeley Laboratory, Building performance database | Department of energy, (n.d.). https://energy.gov/eere/buildings/building-performance-database (Accessed 3 December, 2017). [36] New York Independent System Operator, Power Trends 2016 -The Changing Energy Landscape, (2016), pp. 1–74.

[19] K. Shi, B. Yu, Y. Huang, Y. Hu, B. Yin, Z. Chen, L. Chen, J. Wu, Evaluating the ability of NPP-VIIRS nighttime light data to estimate the gross domestic product and the electric power consumption of China at multiple scales: a comparison with DMSPOLS data, Rem. Sens. 6 (2014) 1705–1724, http://dx.doi.org/10.3390/rs6021705. [20] R. Goldblatt, G. Hanson, K. Heilmann, A. Khandelwal, What's the Matter with Nightlights? Using Landsat Imagery to Improve Nightlight-based Measures of Local Economic Activity with an Application to India, (2016), pp. 1–18. [21] C.N.H. Doll, CIESIN Thematic guide to night-time light remote sensing and its applications, Earth Sci. (2008) 1–41. [22] T. Ma, C. Zhou, T. Pei, S. Haynie, J. Fan, Quantitative estimation of urbanization dynamics using time series of DMSP/OLS nighttime light data: a comparative case study from China's cities, Remote Sens. Environ. 124 (2012) 99–107, http://dx.doi. org/10.1016/j.rse.2012.04.018. [23] H. Letu, M. Hara, H. Yagi, K. Naoki, G. Tana, F. Nishio, O. Shuhei, Estimating energy consumption from night-time DMPS/OLS imagery after correcting for saturation effects, Int. J. Rem. Sens. 31 (2010) 4443–4458, http://dx.doi.org/10.1080/ 01431160903277464. [24] N. Zhao, F. Hsu, G. Cao, E.L. Samson, Improving accuracy of economic estimations with VIIRS DNB image products, Int. J. Rem. Sens. 38 (2017) 5899–5918, http://dx. doi.org/10.1080/01431161.2017.1331060. [25] S. Amaral, G. Câmara, A.M.V. Monteiro, J.A. Quintanilha, C.D. Elvidge, Estimating population and energy consumption in Brazilian Amazonia using DMSP night-time satellite data, Comput. Environ. Urban Syst. 29 (2005) 179–195, http://dx.doi.org/ 10.1016/j.compenvurbsys.2003.09.004. [26] C.C.M. Kyba, S. Garz, H. Kuechly, A.S. de Miguel, J. Zamorano, J. Fischer, F. Hölker, High-resolution imagery of earth at night: new sources, opportunities and challenges, Rem. Sens. 7 (2015) 1–23, http://dx.doi.org/10.3390/rs70100001.

264