We then established actual correlations between the
DN
values
of 2002 stable light annual image composite and the amounts
of CO
2
emissions at the pixel level and produced CO
2
emis-
sions maps using these correlations. Finally, we compared ad-
vantages and disadvantages of the application of the assumed
and actual correlations to map CO
2
emissions, and then
discuss shortcomings of using only nighttime light image data
(rather than further
GIS
manipulation or further use of remote
sensing data) to map CO
2
emissions.
Data
A version-4
DMSP-OLS
stable lights annual image composite
for 2002 was obtained from
NOAA
’s
NGDC
(Earth Observation
Group, 2012). This annual image composite was produced by
all the available cloud-free nighttime light images for that par-
ticular calendar year in the
NGDC
’s digital archive. Each of the
cloud-free nighttime light images were collected by a sensor
with a visible to near-infrared band (0.4 - 1.1 μm) from satel-
lite F15. Ephemeral lights (typically fires) and background
noise have been removed from the annual image composite.
The annual average brightness of the nighttime lights is rep-
resented by 6-bit
DN
s and consequently varies from 0 to 63 in
the image composite. Pixels with
DN
value of 63 are defined
as saturation pixels (Elvidge
et al.
, 2009; Doll, 2008). Detailed
algorithms and processes for the annual image composites
have been described by Baugh
et al.
(2010).
The Vulcan US fossil fuel CO
2
emission inventory for the
year 2002 was used as baseline reference CO
2
emissions data
and obtained from Arizona State University (available from:
). Total Vulcan CO
2
emission is composed of eight sector emissions: residential,
commercial, industrial, power production, on-road mobile,
non-road mobile, aircraft, and cement production sector
emissions. The residential, commercial, and industrial sector
emissions that are from nonpoint data sources are downscaled
by the US Census tract-level demographic data. The on-road
emissions are spatially downscaled by
GIS
road date. Finally,
the total Vulcan emission and each sector emission are raster-
ized to 10 km × 10 km. The Vulcan inventory has been com-
pared to independent estimates at aggregated scales (Gurney
et al.
, 2009). The Vulcan inventory was found to have very
high agreement with CO
2
emissions datasets established by
Department of Energy and the United States Environmental
Protection Agency at the national level and is closer to actual
CO
2
emissions at a spatial resolution of 0.1° × 0.1° than CO
2
emissions datasets produced by population-based approaches
(see page 5539 in Gurney
et al.
, 2009). The largest uncertainty
in the Vulcan inventory is derived from estimations of vehicle
travel miles because vehicle travel miles can directly impact
amounts of CO
2
emissions in on-road and non-road mobile
sectors (Gurney
et al.
, 2009). A detailed description of the
methodology for producing the Vulcan CO
2
emission inven-
tory has been described by Gurney
et al.
(2009).
A vector layer of the contiguous US boundary was taken
from the digital chart of the world by Environmental Systems
Research Institute (Esri) (Denko, 1992).
Methodology
The basic logic for spatially disaggregating or mapping CO
2
emissions is that a region with more commerce and industry
and a larger population usually has brighter lights at night
and emits more CO
2
. However, electric power plants are
point-CO
2
-emissions-sources where a large amount of CO
2
is emitted, but nighttime lights are not bright enough to be
commensurate with the emitted amount of CO
2
. Additionally,
since electric power plants are point- CO
2
-emissions-sources,
it is unnecessary to disaggregate the CO
2
emissions derived
from the power plants. Amounts of CO
2
emitted from power
plants and the power plants’ geographical coordinates can
be obtained from Arizona State University (available from:
), the U.S. Envi-
ronmental Protection Agency (available from:
gov/ghgreporting/ghgdata/index.html
), and Carbon Monitor-
ing for Action (available from:
). In previous
studies using nighttime light image data to disaggregate or
map CO
2
emissions, emissions from electric power plants
were excluded (Ghosh
et al.
, 2010a; Oda and Maksyutov,
2011). Thus, in this study we also subtracted electric power
production sector emissions from the total Vulcan CO
2
emis-
sions as a primary step. Emulating Ghosh
et al.
(2010a) and
Oda and Maksyutov’s (2011) approaches that national CO
2
emissions are disaggregated to each pixel in proportion to
the
DN
values of the nighttime light images, we overlaid the
contiguous US boundary layer on the Vulcan CO
2
emissions
raster layer (Figure 1) to obtain total CO
2
emission (minus CO
2
emissions of power production) in the contiguous US. Again,
we overlaid the contiguous US boundary layer on the stable
lights annual image to gain total sum light (
SL
) that is equal to
the sum of the
DN
values of all the lit pixels in the contiguous
US boundary. Then, the amount of CO
2
emissions indicated
by one unit of sum light (symbolized by U) was calculated.
Finally a CO
2
emissions map (Figure 2a) was produced by
multiplying the stable lights image with the coefficient of U.
The whole process of producing the CO
2
emissions map (Fig-
ure 2a) can be simplified by equation 1:
CO DN
CO
SL
DN
2
2
= × = ×
U
(1)
where
DN
is the
DN
value of the stable lights image.
The above linear function was developed with an assumed
relationship between the
DN
values of stable lights imagery
and the amounts of CO
2
emissions. To explore actual correla-
tion levels between the
DN
values of stable lights imagery and
the amounts of CO
2
emissions, we overlaid the stable lights
annual image composite and the Vulcan CO
2
emissions raster
layer to calculate averaged amounts of CO
2
emissions corre-
sponding to pixels with the same
DN
values of the nighttime
images. Figure 3a shows that an exponential function (Equa-
tion 2) can more accurately describe the relationship between
the
DN
values of the nighttime image and the amounts of CO
2
emissions in the contiguous US.
CO
2
= 78.769e
0.0436×
DN
(2)
A CO
2
emissions map (Figure 2b) of the contiguous US was
produced based on the exponential function in Figure 3a. We
particularly noted that the amount of CO
2
emissions cor-
responding to pixels with a
DN
value of 0 should be 78.77
tonnes C/km
2
/year using the exponential function. In this
analysis we made the assumption, as have previous studies,
that regions without observed nighttime lights are undevel-
oped areas and consequently should not produce any of the
socioeconomic factors being estimated (Ghosh
et al.
, 2010a;
Oda and Maksyutov, 2011; Zhao
et al.
, 2012). We assumed
that pixels with a
DN
value of 0 have 0 tonnes C/km
2
/year of
CO
2
emissions.
Although the exponential correlation is very significant (R
2
= 0.88, P <0.01), the amounts of CO
2
emissions are consider-
ably over-estimated between the
DN
values of 50 to 60 and
greatly under-estimated when the
DN
value is larger than
60 (Figure 3a). A linear function more accurately describes
the correlation when the
DN
values are smaller than 50. Yet,
after the
DN
value reaches 50 an exponential function is the
936
December 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING