Professional Documents
Culture Documents
Estimating Ground-Level PM2.5 Concentration Using Landsat 8 in Chengdu, China
Estimating Ground-Level PM2.5 Concentration Using Landsat 8 in Chengdu, China
Estimating Ground-Level PM2.5 Concentration Using Landsat 8 in Chengdu, China
ABSTRACT: An empirical multilinear model was developed for estimating ground-level PM2.5
concentration at city scale (Chengdu, China) using Landsat 8 data. In this model, the improved DDV (dense
dark vegetation) algorithm (V5.2) was used to retrieve aerosol optical thickness (AOT), Image-based Method
(IBM) was used to compute the land surface temperature (LST), and TVDI was calculated to reflect the air
humidity. The three parameters (AOT, LST, TVDI) and in-situ measured PM2.5 (particulate matter) data
were then utilized to establish the empirical model by partial least square (PLS) regression. In the
computation, the band 9, cirrus band, was used to reduce the influence of atmospheric vapor to LST retrieval.
The results show that when considering moisture and temperature, the correlation between AOT (Aerosol
Optical Thickness) and PM2.5 would be efficiently improved; furthermore, moisture shows more impact on
the relationship than temperature. Station record hourly average PM2.5 also shows higher correlation
coefficients than 24-hr average. As a result, PM2.5 concentration distribution across Chengdu was retrieved
using this model developed in this paper. The method could be a beneficial complement to ground-based
measurement and implicate that remote sensing data has enormous potential to monitor air quality at city
scale.
KEYWORDS: Remote Sensing, Aerosol, PM2.5, Landsat 8, Urban, PLS (Partial Least Square)
1. INTRODUCTION
PM2.5 refers to the particles which is less than or equal to 2.5 micron, also known as inhalable particles. Although
PM2.5 is rarely in the atmospheric, it has an important influence on air quality, visibility, etc. In addition, the size of PM2.5
is very small, so, it can be easily inhaled directly into the bronchi, interface with the gas exchange of the lung, causing
some disease such as asthma, bronchitis and cardiovascular disease, etc. Epidemiologic studies indicates that there are
strong links between PM10 or PM2.5 and public morbidity, mortality of respiratory and cardiovascular disease [1, 2]. Natural
(dust and volcanic ash) and anthropogenic aerosols (biomass burning, smoke, industrial pollution) has gained attention
increasingly, because they can change the property of cloud, and affect the atmospheric circulation of the earth-atmosphere
system by increasing the reflection of solar radiation to space through a variety of complex radiation and microphysical
process [3, 4]. The PMs (PM2.5 and PM10) concentration has become an important index of air pollution and gained more
and more attention from the administrations and organizations of environmental protection and public health all over the
world. Along with the rapid development of economy, PMs has become the primary air pollutants in most major cities of
China, which not only decrease the atmospheric visibility and city scenery, but also threatens people’s health[5].
Air quality monitoring devices are deployed in Chengdu city since 2012, monitoring the air quality (include, the
Remote Sensing of the Atmosphere, Clouds, and Precipitation V, edited by Eastwood Im, Song Yang, Peng Zhang,
Proc. of SPIE Vol. 9259, 925917 · © 2014 SPIE · CCC code: 0277-786X/14/$18 · doi: 10.1117/12.2068886
Table 1. Summary of the models between AOT and PM 2.5 established in previous studies
) %Oa gift'
N914li w :y.ui[4.YS
V
Fig1. The distribution of the monitoring station of Chengdu city (by Google Earth)
Two Landsat 8 images were used in this study; one taken on 03:33:26(UTC), Aug. 13th, 2014, was utilized to construct
the model; another one taken on 03:35:10, Apr. 20th, 2013, was utilized to validate the result. The track numbers of these
two images are 129/39 (path/row). Since the original products are delivered in 16-bit unsigned integer format (DN), Top
Of Atmosphere (TOA) reflectance conversion was carried out to every band before following procession.
2.3 Ground-based PM2.5 data
The in-situ measured PM2.5 data in our study were collected from Chengdu Environment Monitoring Center
(http://www.cdemc.org/pm25.aspx), which reports the air quality condition (include, the concentration of sulfurous acid anhydride,
carbon dioxide, carbon oxide, ozone, PM2.5 and PM10, and air quality index, etc) hourly measured at ground-level. We collected 24-
hr average PM2.5 concentration and the hourly average when the satellite overpass from the seven monitor sites (Table 2).
3. METHODOLOGY
3.1 AOT Retrieval
AOT is retrieved using the Look-up Table (LUT) that contains pre-computed simulations of aerosol type and AOT
value at 0.55µm, which is based on the Second Simulation of the Satellite Signal in the Solar Spectrum (6S) atmosphere
transmission model[29].
In the 6S model, the apparent reflectance, at top of atmosphere (TOA), is a function of successive orders of radiation
interactions, within the coupled surface-atmosphere system. The 6S model assumes that the surface is Lambert surface,
and the atmospheric level is uniform. The atmospheric apparent reflectance (upward reflectance) observed at satellite view
can be described as:
T (θ s ) • T (θν ) • ρ s (θ s , θν , ϕ )
ρtoa (θ s ,θν , ϕ ) = ρ 0 (θ s , θν , ϕ ) + (1)
[1 − ρ s (θ s , θν , ϕ ) • S ]
Where, θs, θv and φ are the solar zenith, the satellite zenith, and the relative azimuth angles(between the sun and satellite);
ρtoa(θs,θv,φ) denote atmospheric apparent reflectance; ρ0(θs,θv,φ) is the reflectance which is equivalent to atmospheric path
radiation; ρs(θs,θv,φ) is the surface reflectance; T(θs) is the normalized downward flux for zero surface reflectance; T(θv)
represents upward total transmission into the satellite field of view; S is the atmospheric backscattering ratios.
DDV algorithm is the widely used algorithm for retrieving aerosol. The algorithm is based on the VIS/SWIR surface
reflectance assumption: over vegetated and dark soiled surfaces (dark pixel), the surface reflectance in some visible
wavelengths correlated with the surface reflectance in the SWIR[30]. Reflectance in 2.2µm band less than 0.15 are usually
considered as dark pixels. Reflectance in 2.2µm band larger than 0.15 are not considered as dark pixels, since the
relationship between the reflectance in the visible channels and the 2.2µm channel is chaotic[31].
Usually, DDV algorithm wasn’t considered to fit to urban area because of its high reflectance. But after analyzing the
imagery of Chengdu city, we found that the reflectance of band 7 (2.2µm) in most area of Chengdu is below 0.15 due to
green mosses and plenty of rooftop garden. In view of this, for most southern China cities with humid climate, the DDV
algorithm could be applied.
ρ sred = ρtoa
2.2
* slopered / 2.2 + y int red / 2.2
(2)
channel, respectively.
The LUT is computed at both red and blue channels using 6S code, and the aerosol model-dependent parameters of eq.
(1) are calculated for a set of AOT value (0.55µm), a variety of geometry conditions, and two aerosol model (continental
and urban). The AOT algorithm in this paper for red and blue channels uses two different LUT, LUTred and LUTblue.
Additionally, the interval (1nm) of Landsat 8’s RSR (relative spectral responses) was resampled to 2.5 nm to fit the spectral
integration step of 6S.
3.2 Land Surface Temperature
Landsat 8 data has thermal infrared band (band 10 and 11) which can be used to retrieve land surface temperature (LST).
Up to now, four main methods are used to retrieve LST: Radioactive Transfer Equation Method (RTEM), Image-based
Method (IBM), Mono-window Algorithm (MWA) and Single-channel Method (SCM). Due to lack of atmospheric
profile when the satellite overpasses, IBM was used to retrieve LST in this study.
Whether IBM, MWA, or SCM, they all need a parameter: land surface emissivity (ε). Different methods were developed
to compute ε [33-35]. In this study, three kinds of land surface features (water surface, town surface and natural surface) were
set: water surface only consists of water body; town surface consists of city and village, mainly include road, all kinds of
construction and buildings; natural surface mainly include all kinds of natural land surface, woodland and farmland, etc.
T maximum likelihood classification mothed was used to divide the whole research area into three kinds of land surface
features.
ρ = ρ * − ρc / Ka (13)
Where, ρ is referred to as reflectance of the “cirrus-path-radiance-corrected” band; ρ* is the original reflectance of the
correcting band; ρc is the reflectance of cirrus band. Ka is the empirical parameter derived from landsat 8 data itself. It is
0.65 in this study.
The NDVI value between vegetation and bare soil indicates that the pixel consists of a certain percentage of vegetation
and a certain percentage of bare soil. So, we determined the vegetation coverage of every pixel in the image with the
following formula:
Where, NDVIv is the NDVI value of vegetation, NDVIs is the NDVI value of bare soil, respectively.
If there are obvious area of dense vegetation in the image, we take the mean NDVI of the dense vegetation area as
NDVIv. Similarly, we take the mean NDVI of the bare soil area as NDVIv. However, if we can’t find the obvious area of
dense vegetation or bare soil, Pv can be calculated with the following assumption, where, NDVIs is 0.05, NDVIv is 0.70.
The vegetation coverage (Pv) of every pixel can be calculated with the following formula:
0 NDVI < NDVIs
Pv = [( N D V I − N D V I s ) / ( N D V I v − N D V I s )]2 ND VIs ≤ ND VI ≤ ND VIv (1 5 )
1 N D VI > N D VIv
In this study, ε of water surface pixels is 0.995, ε of town and natural surface can be calculated with the following
formula:
Where, εsurface is the land surface emissivity of natural surface, and εtown is the land surface emissivity of town surface.
In this study, band10 on Landsat 8 to retrieve the brightness temperature of the land surface, and then calculated the
land surface temperature using ε[36].
The land surface brightness temperature can be calculated with the following formula:
Trad =1282.71/ ln[1+ 666.09/ (3.2 + 0.0370588*DN)] (18)
Where, Trad is the land surface brightness temperature, and DN is DN value of Landsat 8 band10.
Finally, land surface temperature (Tlst) was calculated as:
Trad
Tlst = (19)
1 + (λ * Trad / ρ ) * ln(ε )
ACKNOWLEDGMENT
This research was funded by the prior Research Program of the 12th Five-year Civil Aerospace Plan (D040201-04)
and the science and technology projects of State Grid Corporation (521997140007).
[1] Wallace, L., Correlations of personal exposure to particles with outdoor air measurements: a review of recent studies.
Aerosol Science & Technology, 2000. 32(1): 15-25.
[2] Pope III, C.A., Review: epidemiological basis for particulate air pollution health standards. Aerosol Science &
Technology, 2000. 32(1): 4-14.
[3] Ramanathan, V., P. Crutzen, J. Kiehl, et al., Aerosols, climate, and the hydrological cycle. Science, 2001. 294(5549):
2119-2124.
[4] Kaufman, Y.J., D. Tanré, and O. Boucher, A satellite view of aerosols in the climate system. Nature, 2002. 419(6903):
215-223.
[5] Song, Y., X. Tang, C. Fang, et al., Relationship between the visibility degradation and particle pollution in Beijing.
Acta Scientiae Circumstantiae, 2003. 23(4): 468-471.
[6] Kaufman, Y.J. and R.S. Fraser, Light extinction by aerosols during summer air pollution. Journal of climate and applied
meteorology, 1983. 22(10): 1694-1706.
[7] Chu, D.A., Y. Kaufman, G. Zibordi, et al., Global monitoring of air pollution over land from the Earth Observing
[8] Engel-Cox, J.A., C.H. Holloman, B.W. Coutant, et al., Qualitative and quantitative evaluation of MODIS satellite
sensor data for regional and urban scale air quality. Atmospheric Environment, 2004. 38(16): 2495-2509.
[9] Kacenelenbogen, M., J.-F. Léon, I. Chiapello, et al., Characterization of aerosol pollution events in France using
ground-based and POLDER-2 satellite data. Atmospheric Chemistry and Physics, 2006. 6(12): 4843-4849.
[10] Liu, Y., J.A. Sarnat, V. Kilaru, et al., Estimating ground-level PM2. 5 in the eastern United States using satellite remote
sensing. Environmental science & technology, 2005. 39(9): 3269-3278.
[11] Zhongting, W., L. Qing, T. Jinhua, et al., Monitoring of aerosol optical depth over land surface using CCD camera on
HJ-1 satellite. China Environmental Science, 2009. 29(9): 902-907.
[PM 2.5] = eβ0 *( AOT )βAOT *(TVDI )βALT *(LST )βLST (26)
Where, the [PM2.5] data were collected from ground-based measurement, which are PM2.5 concentration one-hour
average and the 24-hour average when the satellite overpasses. The predictor variables are as follows: the AOT are
retrieved from Landsat 8 data, LST is land surface temperature (LST) retrieved from the thermal infrared band data (band10)
of the Landsat 8 data using Image-based Method (IBM). β0, βAOT, βTVDI and βLST are regression coefficients for AOT, LST,
TVDI and other factors, respectively. For the accuracy, (21) is taken a log-transformation as:
ln[ PM 2.5 ] = β0 + β AOT *ln( AOT + 1) + βTVDI *ln(TVDI + 1) + β LST *ln( LST ) (27)
The natural logarithm transformations can increase the accuracy of the regression coefficients by reducing the influence
of the data concentration of AOT, TVDI and LST. AOT+1 and TVDI+1 are to guarantee their result of ln non-negative.
Partial Least Square method was employed to calculate the regression coefficients.
Table 3 Regression coefficients of different pm2.5 prediction model
PM2.5(µg/m3)
AOT LSH LST (K) R2
hour average 24-hr average
* * 0.5820
* * * 0.6872
* * * 0.7011
* * * * 0.7593
* * * * 0.8279
* * * 0.7505
* * * 0.7222
* * 0.6700
The prediction model was established through Matlab and the procedures are mainly processed with the earth resource
data analysis system, ENVI 4.8. The statistical significance of parameter estimates was reported at α=0.05 level. In the
study, we had constructed the PM2.5 prediction model with different parameter combination (AOT, TVDI and LST) in
logarithm form. Table 3 shows the result of different parameter combinations. The result shown that AOT as the only
predictor of the prediction model has the less effective to explain the PM2.5 concentration than the multiple. Moreover,
compared with PM2.5 (24-hour average), PM2.5 (one-hour average) concentration has a higher correlation. Furthermore,
LST had some impact on the model performance, but the influence is less than other factors. With the purpose of improving
the effective and availability of PM2.5 prediction model, AOT, TVDI and LST are taken as the parameters for PM2.5
prediction model and PM2.5 (hourly average) as the ground-level PM2.5 concentration.
%TIP
104° 0' 0"E
a
:1
°
.. ,
104° 5'0 E
AOT
0-0.27
- 0.27 -0.45
- 0.45 - 0.67
-0.57-0.65
- 0.65 - 0.16
. . - 0.78 - 0.96
104° 10' 0 "E
1
-0.96-1.22
. - 1.22 - 1.61
-1.61-2.17
...'` . N' -2.17-9
104° 0'0"E 104° 5'0"E 104° 10'0 "E
Fig.2 Landsat 8 Chengdu true color image Fig 3 AOT distribution of Chengdu City
(blue=band 2, green=band 3, red=band 4,
Green star: PM2.5 monitoring stations)
z ::
M
'
(1).the retrieval of AOT tends to be more easily influence by ambient conditions, such as dense cloud, water, fog, and
terrain et.al, which may make deviations on the estimation of PMs; (2). the ground-based measurement just only reflect
the PM2.5 concentration at one point, which mainly depend on the local pollutant source and the meteorological features,
rli Ì
i
,i
rir,i
'
ina° n'n"F
,', '
e
D4hi We Roed'
.a;r + 1i-0i
i ,
s
t
4!
t'
1
'
- u.'
IC
104° 5'0"E
`1, `
.111r
lM 11'jte,,i 1//..
7
17.Ij'
luninuSrreet
I
e
.
4
while the satellite estimation represents the average situation within a pixel, and thus the information they obtained may
be very different; (3) both LST and TVDI used for correct the relationship between AOT and PM2.5 are also average at
one pixel, which can’t represent the every point region.
,
1
1 z
ò
PM2.5 (µm/m3)
- 9.3- 17.7
- 17.7 -22.4
- 22.4- 25.04
- 25.0 -26.5
- 26.5 -29.2
- 29.2-33.9
- 33.9 -42.3
- 42.3 - 57.2
-57.2 -84.0
- 84.0 -131.8
Fig.4. PM2.5 distribution. Left: PM2.5 distribution over Chengdu city; Right: PM2.5 distribution over Chengdu with the reference of road net
This paper proposes a multilinear model of PM2.5 concentration prediction using Landsat 8 data, the results show that
the model has a higher correlation coefficient (R2=0.7451). When considering TVDI and temperature, the correlation
between AOT and PM2.5 would be significantly improved; furthermore, TVDI shows larger impact on the relationship
than temperature, and using PM2.5 hourly average had higher correlation coefficients than using PM2.5 24-hr average.
Due to lack of air monitor sites, the model performance cannot evaluate effectively, but only compared retrieved PM2.5
concentration with PM2.5 ground-based measurement. Although log-transformation modifies the non-normality of PM2.5
data and gives more accurate standard errors of the regression coefficients, it constrains the relationship between PM2.5
concentration and each of these predictors to an exponential form. So, the model still needs to further modify for applying
in the future. There are some epidemiologic studies of particle which usually under limitation of ground-based
ACKNOWLEDGMENT
This research was funded by the prior Research Program of the 12th Five-year Civil Aerospace Plan (D040201-04)
and the science and technology projects of State Grid Corporation (521997140007).
[1] Wallace, L., Correlations of personal exposure to particles with outdoor air measurements: a review of recent studies.
Aerosol Science & Technology, 2000. 32(1): 15-25.
[2] Pope III, C.A., Review: epidemiological basis for particulate air pollution health standards. Aerosol Science &
Technology, 2000. 32(1): 4-14.
[3] Ramanathan, V., P. Crutzen, J. Kiehl, et al., Aerosols, climate, and the hydrological cycle. Science, 2001. 294(5549):
2119-2124.
[4] Kaufman, Y.J., D. Tanré, and O. Boucher, A satellite view of aerosols in the climate system. Nature, 2002. 419(6903):
215-223.
[5] Song, Y., X. Tang, C. Fang, et al., Relationship between the visibility degradation and particle pollution in Beijing.
Acta Scientiae Circumstantiae, 2003. 23(4): 468-471.
[6] Kaufman, Y.J. and R.S. Fraser, Light extinction by aerosols during summer air pollution. Journal of climate and applied
meteorology, 1983. 22(10): 1694-1706.
[7] Chu, D.A., Y. Kaufman, G. Zibordi, et al., Global monitoring of air pollution over land from the Earth Observing
[8] Engel-Cox, J.A., C.H. Holloman, B.W. Coutant, et al., Qualitative and quantitative evaluation of MODIS satellite
sensor data for regional and urban scale air quality. Atmospheric Environment, 2004. 38(16): 2495-2509.
[9] Kacenelenbogen, M., J.-F. Léon, I. Chiapello, et al., Characterization of aerosol pollution events in France using
ground-based and POLDER-2 satellite data. Atmospheric Chemistry and Physics, 2006. 6(12): 4843-4849.
[10] Liu, Y., J.A. Sarnat, V. Kilaru, et al., Estimating ground-level PM2. 5 in the eastern United States using satellite remote
sensing. Environmental science & technology, 2005. 39(9): 3269-3278.
[11] Zhongting, W., L. Qing, T. Jinhua, et al., Monitoring of aerosol optical depth over land surface using CCD camera on
HJ-1 satellite. China Environmental Science, 2009. 29(9): 902-907.
[17] Wang, J. and S.A. Christopher, Intercomparison between satellite‐derived aerosol optical thickness and PM2. 5 mass:
Implications for air quality studies. Geophysical research letters, 2003. 30(21).
[18] Péré, J.-C., V. Pont, M. Mallet, et al., Mapping of PM10 surface concentrations derived from satellite observations of
aerosol optical thickness over South-Eastern France. Atmospheric Research, 2009. 91(1): 1-8.
[19] Vidot, J., R. Santer, and D. Ramon, Atmospheric particulate matter (PM) estimation from SeaWiFS imagery. Remote
sensing of environment, 2007. 111(1): 1-10.
[20] Pelletier, B., R. Santer, and J. Vidot, Retrieving of particulate matter from optical measurements: a semiparametric
approach. Journal of Geophysical Research: Atmospheres (1984–2012), 2007. 112(D6).
[21] Wang, Z., L. Chen, J. Tao, et al., Satellite-based estimation of regional particulate matter (PM) in Beijing using
vertical-and-RH correcting method. Remote sensing of environment, 2010. 114(1): 50-63.
[22] Liu, Y., R.J. Park, D.J. Jacob, et al., Mapping annual mean ground‐level PM2. 5 concentrations using Multiangle
Imaging Spectroradiometer aerosol optical thickness over the contiguous United States. Journal of Geophysical
Research: Atmospheres (1984–2012), 2004. 109(D22).
[23] van Donkelaar, A., R.V. Martin, M. Brauer, et al., Global estimates of ambient fine particulate matter concentrations
from satellite-based aerosol optical depth: development and application. Environmental Health Perspectives,
2010. 118(6): 847.
[24] Gupta, P., S.A. Christopher, J. Wang, et al., Satellite remote sensing of particulate matter and air quality assessment
over global cities. Atmospheric Environment, 2006. 40(30): 5880-5892.
[25] Hänel, G., The properties of atmospheric aerosol particles as functions of the relative humidity at thermodynamic
equilibrium with the surrounding moist air. Advances in geophysics, 1976. 19: 73-188.
[26] Kotchenruther, R.A. and P.V. Hobbs, Humidification factors of aerosols from biomass burning in Brazil. Journal of
Geophysical Research: Atmospheres (1984–2012), 1998. 103(D24): 32081-32089.
[27] Raut, J.-C. and P. Chazette, Retrieval of aerosol complex refractive index from a synergy between lidar, sunphotometer
and in situ measurements during LISAIR experiment. Atmospheric Chemistry and Physics, 2007. 7(11): 2797-
2815.