Spectral, spatial, and temporal sensitivity ofcorrelating MODIS aerosol optical depth withground-based fine particulate matter (PM2.5)

across southern Ontario

Jie Tian and Dongmei Chen

Abstract. This paper evaluates the sensitivity of the aerosol optical depth (AOD) measurements derived from the Moderate

Resolution Imaging Spectroradiometer (MODIS) at wavelengths of 0.47, 0.55, and 0.66 mm from both the Terra and Aqua

satellites for the estimation of ground-level concentrations of fine particulate matter (PM2.5) in southern Ontario, Canada.

The correlation between MODIS AOD and ground-based PM2.5 measurements is compared at different seasons and

spatial aggregation levels. The results showed that the MODIS AOD data acquired in early afternoon (from Aqua)

seem to have slightly higher agreement with the ground-based measurement of PM2.5 than the late morning Terra data.

Moreover, MODIS AOD at 0.47 mm from both satellites had the highest correlation with ground-level PM2.5. More

detailed examination suggested that the correlation was stronger in the spring and summer and weaker in the fall and

winter. Aqua MODIS AOD appeared to have a better correlation with the average ground-based PM2.5 concentration over

1–3 h than with daily PM2.5. MODIS AOD values aggregated over 3 6 3 pixel groups correlate slightly better with PM2.5

than with the original single centre pixel values.

Resume. Dans cet article, on evalue la sensibilite des mesures de l’epaisseur optique des aerosols (AOD) derivees du capteur

MODIS (« Moderate Resolution Imaging Spectroradiometer ») dans les longueurs d’onde de 0,47 mm, 0,55 mm et 0,66 mm

des satellites Terra et Aqua pour l’estimation de la concentration au niveau du sol des particules fines (P2,5) dans le sud de

l’Ontario, au Canada. Les correlations entre l’AOD de MODIS et les mesures au sol de P2,5 sont comparees a differentes

saisons et pour differents niveaux d’agregation spatiale. Les resultats ont demontre que les donnees d’AOD de MODIS

acquises tot en apres-midi (a partir d’Aqua) semblent afficher une correlation legerement plus grande avec la mesure au sol

de P2,5 que les donnees acquises en fin d’avant-midi par Terra. En outre, l’AOD de MODIS a 0,47 mm, dans le cas des deux

satellites, etait meilleure dans la correlation globale par rapport aux mesures de P2,5 au niveau du sol. Un examen plus

detaille a suggere que la correlation etait plus forte au printemps et en ete et plus faible a l’automne et en hiver. L’AOD de

MODIS acquise par Aqua semblait avoir une meilleure correlation avec la concentration moyenne de P2,5 au sol sur 1–

3 heures que la mesure journaliere de P2,5. Les valeurs d’AOD de MODIS agregees par rapport a des groupes de pixels de

3 6 3 donnent une correlation legerement superieure avec les valeurs de P2,5 par comparaison avec les valeurs originales des

pixels a centre unique.

[Traduit par la Redaction]

Introduction

Particulate matter (PM) is one of the major pollutants

affecting regional air quality and is commonly recognized

as a complex mixture of microscopic solid or liquid particles

suspended in the air. PM is usually characterized according

to an aerodynamic size because of the different health effects

that develop from exposure to PM of various diameters. In

particular, fine PM (PM2.5) often refers to particles that are

2.5 mm or less in diameter. PM2.5 originates from a variety of

sources ranging from natural production (e.g., volcanic

eruptions and forest fires) to anthropogenic pollution (e.g.,

emissions from industrial facilities and combustion from

automobiles) (Arya, 1998). PM2.5 usually remains suspended

in the air for relatively long periods of time as compared with

coarser PM and can penetrate deep into the respiratory sys-

tem and cause disease and even premature death (Villeneuve

et al., 2002). Children, the elderly, and those with asthma,

respiratory problems, or cardiovascular or lung disease are

found to be most susceptible to PM2.5 (Pope et al., 1999;

Green and Armstrong, 2003; Kappos et al., 2004; Neuberger

et al., 2004).

A common practice in many industrialized countries to

monitor ambient air quality regionally is to establish a net-

m10-033.3d 27/8/10 19:12:57

Proof/Epreuve

Received 2 November 2009. Accepted 18 May 2009. Published on the Web at http://pubservices.nrc-cnrc.ca/cjrs on XXXX XXXX XXXX.

J. Tian and D. Chen.1 Department of Geography, Queen’s University, Kingston, ON K7L 3N6, Canada.

1Corresponding author (e-mail: chendm@queensu.ca).

Can. J. Remote Sensing, Vol. 36, No. 2, pp. 000–000, 2010

Pagination not final/Pagination non finale

E 2010 CASI 1

work of monitoring stations, equipped by receptors, across a

region of interest. However, although having a high tem-

poral sampling frequency, these stations usually take point

measurements and do not provide adequate coverage to ful-

fill the needs of mapping regional air quality. The spatial

interpolation of such station data therefore possesses limited

value to understanding the spatial–temporal dynamics of a

pollutant (Kumar et al., 2007).

Remote sensing provides an alternative data resource that

is important in studying the air quality of a relatively large

area. At the present time, a number of orbiting satellites

routinely measure the spectral radiation received at the top

of the atmosphere and provide data regarding atmosphere

composition. The Moderate Resolution Imaging Spectrora-

diometer (MODIS) is one of the key sensors onboard the

Terra and Aqua satellite platforms and collects data in 36

channels from a near-Earth Sun-synchronous polar orbit

with a revisit frequency of 1–2 days depending on the loca-

tion on the globe (Barnes et al., 1998). Aerosol optical depth

(AOD) is a parameter that measures the integral of atmo-

spheric aerosol extinction from the Earth’s surface to the top

of the atmosphere (Gupta et al., 2006). The MODIS AOD

data are derived at three wavelengths, namely 0.47, 0.55, and

0.66 mm (Ichoku et al., 2004).

Efforts have been made to correlate MODIS AOD and

ground-level PM2.5 concentration at the continental scale

(Donkelaar et al., 2006), the regional scale (Hutchison,

2003; Engel-Cox et al., 2004; Tian and Chen, 2007), and even

the local scale, focusing on a group of individual cities (Chu

et al., 2003; Li et al., 2009). Although MODIS AOD has

been commonly reported to have tremendous potential in

estimating PM2.5, unstable results have been observed in

terms of the strength of the correlation and its spatial–tem-

poral variation. For example, Wang and Christopher (2003)

reported relatively strong correlations (r 5 0.7) for their

study area of Jefferson County, Alabama. Al-Saadi et al.

(2005) and Engel-Cox et al. (2004), however, found weaker

correlations (r , 0.4) in the western US, where the higher

surface reflectance of the more arid surfaces was suspected to

reduce contrast. Another reason for the weaker correlation is

that the AOD retrieval model used for that region assumes

more dust and smoke than industrial aerosols. Furthermore,

air pollution mechanisms can vary spatially, and hence so

also can the strength of the AOD–PM2.5 correlation (Kumar

et al., 2007). Several research projects have been dedicated to

a search for the optimal scale to achieve the best correlation

(Li et al., 2009). However, there have been few studies focus-

ing on the AOD–PM2.5 correlation or its variation at the

regional scale in Canada. Tian and Chen (2007) attempted

to correlate MODIS AOD and PM2.5 in Ontario based on a

relatively small volume of data only from the Terra satellite

and found that there was a positive correlation between

them; the strengths of the correlation appeared to be at dif-

ferent levels in the summer and winter. However, Tian and

Chen did not provide a thorough or systematic analysis of

the correlation variation in terms of spectral, spatial, and

temporal sensitivity and did not compare AOD–PM2.5 cor-

relations from the two satellite platforms, namely Terra and

Aqua (acquiring data late morning and early afternoon local

time, respectively).

This paper presents the results of an inclusive and detailed

study to systematically evaluate the correlation of MODIS

AOD and PM2.5 in southern Ontario, Canada. The MODIS

algorithm for AOD retrieval is first reviewed to highlight the

evolution history of the MODIS AOD data. The MODIS

AOD measurements derived at wavelengths of 0.47, 0.55,

and 0.66 mm are compared in terms of their correlation

strength with PM2.5. The original AOD data are spatially

averaged over 3 6 3 pixel groups, and the hourly PM2.5 data

are temporally aggregated to represent longer time periods.

Detailed cross-comparison is then made to illustrate the

sensitivity of the AOD–PM2.5 correlation to a series of

changes in data resolution. The analyses are performed in

parallel for the AOD data from both Terra and Aqua so that

a sound discussion can be carried out to address possible

diurnal differences. Lastly, the distributional pattern of the

yearly PM2.5 average is produced by interpolating the mon-

itoring station data and is then compared with that of yearly

Terra MODIS AOD.

Review of MODIS algorithm for aerosoloptical depth retrieval

The MODIS algorithm for AOD retrieval consists of two

independent algorithms, one for deriving aerosols over land

and the other for deriving aerosols over oceans. Principally,

the land algorithm is applied only to a subset of those

MODIS pixels that have no ocean or inland water within

the field of view. The AOD retrieval analyses only include

those pixels that are not contaminated by cloud. The core

algorithm uses the atmospheric reflectance ratios between

0.47 and 2.12 mm and between 0.66 and 2.12 mm to retrieve

the AOD information (Remer et al., 2006). The AOD

retrieval is conducted based on a radiative transfer model

and uses Mie theory for fine aerosol models. The algorithms

of various versions differ in their logic of determining which

500 m pixels in a 20 6 20 pixel group should be used to

generate AOD values.

MODIS AOD data are organized by collections, where

groups of data were processed by similar, but not necessarily

the same, versions of the algorithm. The first globally vali-

dated data were produced using the collection 3 algorithms.

Collections 4 and 5 algorithms have improved in the detec-

tion of cloudy pixels. The current collection 5 algorithm over

land (C005-L) is regarded as a complete overhaul to the

collection 4 algorithm over land (C004-L) (Levy et al.,

2007). Whereas C004-L essentially retrieves aerosol prop-

erties independently in two visible channels (0.47 and

0.66 mm), C005-L performs the same operation in three

channels simultaneously (the two visible channels and the

2.12 mm channel). The C005-L algorithm is also superior in

m10-033.3d 27/8/10 19:12:57

Proof/Epreuve

Vol. 36, No. 2, April/avril 2010

Pagination not final/Pagination non finale

2 E 2010 CASI

that it assumes the 2.12 mm channel contains information

about not only the surface but also the coarse mode aerosol.

There are additional differences in surface reflectance

assumptions and the aerosol look-up tables. A more detailed

description of the algorithms can be found in Kaufman et al.

(1997) and Remer et al. (2006).

Study area and ground-based data

This paper focuses on southern Ontario, which is the most

populous region in Canada. Southern Ontario is mainly

within the Great Lakes – St. Lawrence Lowlands physio-

graphic region (Bone, 2005). A continental climate affects

this temperate mid-latitude region. The climate is highly

modified by the influence of the Great Lakes: the addition

of moisture from the lakes increases precipitation amounts.

The latest studies indicate that areas of southern Ontario

often experience the highest levels of PM2.5 in eastern

Canada. Ontario is burdened with Can$9.6 billion in health

and environmental damage each year due to the impact of

ground-level fine PM and ozone (Yap et al., 2005), which are

generally attributed to the formation of smog.

The Ontario Ministry of the Environment (OME) oper-

ates and maintains the air quality monitoring network in

Ontario, which monitors concentrations of particulate and

gaseous air pollutants at the near-ground level (typically 3–

8 m above the ground). More specifically, the ThermoScien-

tific tapered element oscillating microbalance (TEOM) tech-

nology is used to measure ground-level PM2.5 concentrations.

Each station is equipped with a filter attached to a hollow,

tapered, and oscillating glass rod. The accumulation of mass

is measured over time from the change in the oscillation fre-

quency. The PM2.5 readings are provided on an hourly basis

in the standard units of micrograms per cubic metre (mg/m3),

with a precision of about ¡1.5 mg/m3. The PM2.5 data for this

study were obtained from the OME electronic archive of

historical air quality data. In 2004, PM2.5 was routinely mea-

sured by 38 monitoring stations mainly distributed across the

southern part of the province. Figure 1 shows the locations of

the monitoring stations.

Satellite data

MODIS aerosol images were collected for southern

Ontario to cover the entire calendar year of 2004. Collection

5 aerosol images were chosen because they are produced

based on the most advanced version of the retrieval algo-

rithm. Preliminary validation of the algorithm used for the

AOD retrieval in this collection showed much improved

results, where the MODIS – Aerosol Robotic Network (at

0.55 mm) regression has an equation y 5 1.01x + 0.03, with a

correlation coefficient of 0.90 (Levy et al., 2007).

In total, over 800 AOD images were obtained from both

Terra and Aqua. A MODIS aerosol image covers a 5 min

time interval and has typical dimensions of 203 cells (along

swath) 6 135 cells (across swath). The typical cell size, or

spatial resolution, is 10 km 6 10 km (at nadir). For southern

Ontario, there is usually more than one image, taken at dif-

ferent local times, available for each day.

The MODIS aerosol data are stored and provided in a

hierarchical data format (HDF) (Masuoka et al., 1998),

which is a multi-object file format for sharing scientific data

m10-033.3d 27/8/10 19:12:58

Proof/Epreuve

Figure 1. Locations (solid triangles) of the air quality monitoring stations across Ontario.

Canadian Journal of Remote Sensing / Journal canadien de teledetection

Pagination not final/Pagination non finale

E 2010 CASI 3

in multiplatform distributed environments. Each MODIS

aerosol product file stores 53 scientific parameters. Speci-

fically, MODIS AOD includes three channels, with each

containing the AOD values for one wavelength (0.47, 0.55,

or 0.66 mm). A sample image of Terra MODIS AOD (at

0.47 mm) is shown in Figure 2. A valid AOD value is dimen-

sionless and has a valid range from 0 to 5, with values over

unity generally being classified as heavy haze (Engel-Cox et

al., 2004). MATLAB programs were written to read the

MODIS HDF files systematically and extract the AOD

channels for land, which contain the data needed for this

study.

Analysis methods

The same analyses were performed in parallel with the

collected MODIS AOD images from Aqua and Terra for

southern Ontario. Initial attempts focused on correlating

the original MODIS AOD values (single pixels) at the three

wavelengths of 0.47, 0.55, and 0.66 mm, with the hourly read-

ings of ground-level PM2.5 concentration from the ground

stations. The MODIS AOD values and the PM2.5 readings

were associated based on their spatial and temporal prox-

imity to perform a series of intercomparative and quantita-

tive analyses. Each monitoring station was spatially

colocated to its nearest MODIS pixel. MATLAB programs

were developed to search for the nearest pixel from each indi-

vidual MODIS AOD image for every ground station. If the

MODIS pixel found nearest, or more normally, covering the

station had a valid AOD value, this value was associated

with the station. The AOD value was then paired with the

coincident PM2.5 reading that was measured by the asso-

ciated station and within the same hour as the MODIS

satellite overpass. Pearson’s r was calculated to quantify

the strength of the correlation between the hourly PM2.5

and the corresponding MODIS AOD values at the three

wavelengths.

It was borne in mind that PM2.5 and MODIS AOD mea-

sure different atmospheric loadings of aerosols. The former

represents the fine particulate matter concentration at a

point station near the ground, whereas the latter quantifies

the total columnar aerosol loading over a typical area of

10 km 6 10 km. In addition, there is a considerable discrep-

ancy in the temporal characteristics between the two sets of

data. The original PM2.5 data characterize hourly concentra-

tion averages, whereas MODIS AOD is retrieved based on

nearly instant observations. The original MODIS AOD

images were therefore further processed to assign each single

pixel with the average of the valid AOD values within its 3 63 pixel neighborhood. On the other hand, the hourly PM2.5

data were temporally aggregated up to multihour averages

and the daily average as well. A series of correlation analyses

was then performed to test the sensitivity of the AOD–PM2.5

correlation to these data resolution changes.

The overall analyses tend to even off the correlation var-

iations across the region and over time. More detailed exam-

ination was performed to focus on the individual station

sites. It is understandable that AOD and PM2.5 may be

strongly correlated in some geographical areas but weakly

correlated in others due to differences in meteorological

m10-033.3d 27/8/10 19:12:58

Proof/Epreuve

Figure 2. Sample image of MODIS AOD at 0.47 mm covering southern Ontario (acquired on

23 September 2004 and processed by the version 5 algorithm).

Vol. 36, No. 2, April/avril 2010

Pagination not final/Pagination non finale

4 E 2010 CASI

condition (e.g., humidity or atmospheric layering) and (or)

land use – land cover characteristics (e.g., surface reflectiv-

ity). The spatial variation of aerosol type may also account

for the spatial variation of AOD–PM2.5 correlation. More-

over, the AOD–PM2.5 correlation for various months was

examined to elucidate seasonal patterns. To provide some

idea of the data availability for the cities in southern Ontario,

the frequency (or availability) of the valid MODIS AOD

pixel values representing the stations within each month

was summarized.

Spatial interpolation is the most common approach used

to estimate values for unmeasured locations based on some

known point measurements. The yearly mean of PM2.5 was

calculated for each monitoring station and was spatially

interpolated to produce a concentration surface. The inverse

distance weighting (IDW) method was adopted in the inter-

polation process for the sake of simplicity and ease of com-

parison. A yearly AOD average map was produced by

assigning the average values from the Terra MODIS AOD

pixels. Then the interpolated PM2.5 surface map was com-

pared with the yearly AOD average map.

Results and discussion

Ground-level PM2.5 concentrations are classified into five

categories according to the US Environmental Protection

Agency Air Quality Index. These categories were used as a

reference in this study. The air quality category is considered

good if daily PM2.5 concentrations are between 0 and 15.4 mg/

m3, moderate if daily PM2.5 concentrations are between 15.5

and 40.4 mg/m3, unhealthy for sensitive groups (e.g., children

and seniors) if daily PM2.5 concentrations are between 40.5

and 65.4 mg/m3, unhealthy if daily PM2.5 concentrations are

between 65.5 and 150.4 mg/m3, and very unhealthy if daily

PM2.5 concentrations are between 150.5 and 250.4 mg/m3

(USEPA, 2003). Figure 3 uses two cities (Toronto and

Kingston) as an example to illustrate how the daily PM2.5

average, measured by the monitoring stations, changes

throughout the year 2004. As can be seen from Figure 3,

spring and summer exhibit relatively higher levels of PM2.5

and are likely to be more harmful to certain patients, infants,

or seniors compared to fall and winter. It is worth mention-

ing that PM2.5 is concentrated at levels higher than the daily

average during certain hours (e.g., rush hour).

The overall correlation analysis shows good agreement

between the MODIS AOD and the ground-based PM2.5 con-

centration measurements (see Figure 4). Larger AOD values

usually correspond to higher PM2.5 values. More detailed

information about the correlations is provided in Table 1,

which shows that all the correlations are statistically signifi-

cant because the probability values associated with the null

hypothesis are extremely small (,0.001). Table 1 shows how

MODIS AODs are correlated with hourly PM2.5, with coef-

ficients ranging from 0.51 to 0.65. In comparison, hourly

PM2.5 is better correlated with Aqua AODs (r 5 0.649,

0.614, and 0.566 at the three wavelengths) than with Terra

AODs (r 5 0.600, 0.561, and 0.509 at the three wavelengths)

by a slight margin. One possible reason might be that Aqua

MODIS images are taken in the early afternoon when the

temperature becomes relatively higher in the diurnal cycle

and the planetary boundary layer (PBL) is better mixed

due to the stronger convection. The aerosol vertical profile

therefore tends to be even smoother (Raj et al., 2008), and

measurements of columnar aerosol loading (e.g., AOD)

taken at that time are perhaps more representative of the

corresponding PM2.5.

In southern Ontario, MODIS AOD values at 0.47 mm

(mean AOD 5 0.289) are larger than those at 0.55 and

0.66 mm (mean AOD 5 0.236 and 0.197, respectively). A

clear trend is recognized in that the AOD at shorter wave-

lengths has a stronger correlation with PM2.5. The trend

appears to be rather consistent no matter how the PM2.5 data

were scaled up in time. This may be explained by the fact that

the smaller wavelength is more sensitive to smaller particles

and less sensitive to coarse mode. In other words, southern

Ontario is dominated by smaller aerosol particles, and the

MODIS AOD data at 0.47 mm (compared with those at 0.55

and 0.66 mm) are therefore relatively more representative of

the PM2.5.

The 3 6 3 AOD averages at the three wavelengths are also

found to correlate with PM2.5 in a positive manner (see

Figure 5). Overall, PM2.5 is slightly better correlated with

the 3 6 3 AOD averages (r 5 0.53–0.68) than with the single,

centre pixel AOD data (r 5 0.50–0.65). Some of the differ-

ences in correlation strength are significant. For example,

when the Aqua MODIS AOD data are aggregated over

3 6 3 pixel groups, the correlation with hourly PM2.5

increases from 0.649 (n 5 1779) to 0.682 (n 5 3305). Inter-

comparison between the 3 6 3 AOD averages at different

wavelengths shows that PM2.5 correlates best with the 3 6 3

AOD average at 0.47 mm (see Table 2). This somewhat dis-

agrees with earlier published studies (Hutchison et al., 2005),

where weaker correlations were found between PM2.5 and

the AOD average within a 3 6 3 pixel group. Although the

reasons for this disagreement are not yet clear, it may be

because of differences in the land surface characteristics

and changes in the current AOD retrieval algorithm over

the previous algorithms. The spatial averaging over 3 63 pixel groups produces many more AOD values than pro-

duced by the original single centre pixels. This processing

may compensate, to some degree, for scatter in the MODIS

AOD data due to, for example, cloud contamination.

Figure 6 exhibits the correlations between MODIS AOD

at 0.47 mm (chosen as an example) and the PM2.5 data at

various temporal resolutions (averaged over different

lengths of time to a maximum of 24 h). The degree of cor-

relation for Terra MODIS AOD does not seem to be very

sensitive to the temporal resolution of the PM2.5 data. In

comparison, the degree of correlation for Aqua MODIS

AOD appears to decline consistently as the PM2.5 data are

aggregated over increasingly longer intervals of time. Aqua

MODIS AOD is most representative of the coincident

m10-033.3d 27/8/10 19:13:00

Proof/Epreuve

Canadian Journal of Remote Sensing / Journal canadien de teledetection

Pagination not final/Pagination non finale

E 2010 CASI 5

hourly or 3 h PM2.5 than the daily average. Overall, the

strongest correlation (0.685) is between the 3 6 3 AOD

average data from Aqua and the 3 h PM2.5 (see Table 2).

There may be an optimal match between the temporal reso-

lution of the point PM2.5 data and the spatial resolution of

the areal AOD data for an efficient correlation. Analysis

covering a larger spatial extent may allow spatial aggrega-

tions over larger windows and provide further information

related to the optimal spatial resolution.

The availability of MODIS AOD data and their correla-

tion with PM2.5 vary greatly with the seasons (see Figure 7).

For the monitoring station locations, there are few valid

MODIS AOD data available for January, February, March,

and December due to the extensive snow cover (having high

reflectance unfavorable to MODIS AOD retrieval). MODIS

therefore has limited value for PM2.5 estimation for southern

Ontario during winter. There are considerably more valid

AOD data for the study area during the rest of the year.

m10-033.3d 27/8/10 19:13:00

Proof/Epreuve

Figure 3. Variation of the daily ground-level PM2.5 concentration measured by the monitor-

ing stations at Kingston (a) and Toronto (b), Ontario, for 2004.

Vol. 36, No. 2, April/avril 2010

Pagination not final/Pagination non finale

6 E 2010 CASI

The dramatic increase in MODIS AOD data frequency in

April is perhaps because the snow has begun to melt and the

algorithm for MODIS AOD retrieval becomes applicable to

more areas. Somewhat fewer AOD data are available for

August compared with any other summer month, possibly

because the MODIS images acquired during that period are

contaminated by more clouds. Cloud cells need to be

screened out during the AOD retrieval, and too many clouds

will cause failure of retrieving a statistically valid value for

some areas. The significantly greater number of valid AOD

data in the early fall is attributed to the frequent presence of

ideal weather (e.g., clear sky) and land cover conditions (e.g.,

no snow) for MODIS AOD retrieval.

As can be seen in Figures 7a and 7b, the correlations

between MODIS AOD and PM2.5 change significantly from

month to month. May, July, August, and September show

the highest correlation coefficients of 0.70–0.85, and June has

weaker correlations of 0.50–0.60 compared with those of the

other summer months. The period of May–September can be

regarded as optimal for PM2.5 estimation using MODIS

AOD. Despite seasonal changes, MODIS AOD at a shorter

wavelength normally has a stronger correlation with PM2.5.

The estimated distribution of yearly ground-level PM2.5

concentration average is depicted in Figure 8a. Since the

interpolation was performed based only on the monitoring

stations currently available in southern Ontario, the result

appears to be rather general and does not provide much

detailed information on the change of PM2.5 over relatively

short distances. Also, the estimated distribution reflects the

instrumental character of the interpolation method and is

heavily dependent on the number and structure of the mon-

itoring stations. In contrast, the distribution of yearly Terra

MODIS AOD at 0.47 mm offers a much more detailed pic-

ture of aerosol loading across the region (Figure 8b).

Although MODIS AOD is retrieved based on instant obser-

vation and is not a direct proxy of daily PM2.5, a good agree-

ment (Pearson’s r of around 0.60) has been found between

these two variables (see Figure 6). The distribution seems to

have a pattern similar to that of the densely populated and

(or) industrialized areas in the region. Overall, the land

boundary areas are generally loaded with more aerosols than

the inland areas. The Greater Toronto Area, the most popu-

lated region in the province, is shaded in red in Figure 8b,

implying a heavily loaded subregion that is not suggested by

Figure 8a. Relatively heavy aerosol loadings are also found

over the Greater Windsor Area, the southern part of the

Golden Horseshoe, and the region between Grand Bend

and Tiverton. Due to the strong correlation between

MODIS AOD and PM2.5, a higher value of yearly MODIS

AOD can be interpreted as an indicator of potentially higher

concentrations of PM2.5 near the ground. This information

would provide a better regional PM2.5 distribution than

PM2.5 values directly interpolated from monitoring stations.

m10-033.3d 27/8/10 19:13:00

Proof/Epreuve

Figure 4. Scatterplots of hourly ground-level PM2.5 concentration versus MODIS AOD at 0.47 mm from Terra (a) and Aqua (b).

Table 1. Overall correlations between MODIS single-pixel AODs

and hourly, 3 h, and daily ground-level PM2.5 concentrations.

PM2.5

average

AOD

wavelength (mm)

Terra Aqua

r n R n

Hourly 0.47 0.600 2104 0.649 1779

0.55 0.561 2105 0.614 1780

0.66 0.509 2106 0.566 1778

Three hour 0.47 0.608 2126 0.651 1794

0.55 0.566 2127 0.617 1795

0.66 0.513 2128 0.570 1793

Daily 0.47 0.593 2136 0.597 1804

0.55 0.553 2137 0.564 1805

0.66 0.501 2138 0.518 1803

Note: All r values (correlation coefficient) are significant at p , 0.001.n, number of valid pairs.

Canadian Journal of Remote Sensing / Journal canadien de teledetection

Pagination not final/Pagination non finale

E 2010 CASI 7

Conclusions

This paper evaluates the spatial and temporal sensitivity

of the aerosol optical depth (AOD) measurements derived

from the Moderate Resolution Imaging Spectroradiometer

(MODIS) at wavelengths of 0.47, 0.55, and 0.66 mm from

both the Terra and Aqua satellites for the estimation of

ground-level concentrations of fine particulate matter

(PM2.5) in southern Ontario, Canada. As demonstrated in

previous research (Tian and Chen, 2007), overall, MODIS

AOD agrees well with ground-level PM2.5 concentrations.

However, PM2.5 is best correlated with the MODIS AOD

at a wavelength of 0.47 mm compared with at 0.55 or 0.66 mm.

The MODIS AOD acquired in the early afternoon (from

Aqua) seems able to better represent the condition of

PM2.5 than the late morning acquisition from Terra. The

correlation between Terra MODIS AOD and PM2.5 is not

very sensitive to the temporal resolution changes of the

PM2.5 data, but the Aqua MODIS AOD seems to be more

representative of the coincident PM2.5 averaged over 1–3 h

than the daily average. The mean AOD over 3 6 3 pixel

groups tends to have a relatively stronger correlation with

PM2.5 than the original single centre pixel AOD. In practice,

MODIS AOD images often appear to be patchy and can

provide little help in PM2.5 estimation in winter because of

the very limited number of valid data. Moreover, the

MODIS AOD–PM2.5 correlation is significantly stronger

in summer months than in winter months.

The correlation between MODIS AOD and PM2.5 also

varies substantially across space and through time. It is

unclear why the correlation between MODIS AOD and

PM2.5 varies at different spatial and temporal scales. The

potential impact factors would include meteorological con-

ditions, land use, cloud contamination, and station location.

m10-033.3d 27/8/10 19:13:01

Proof/Epreuve

Figure 5. Scatterplots of hourly ground-level PM2.5 concentration versus MODIS 3 6 3 AOD at 0.47 mm from Terra (a) and Aqua (b).

Table 2. Overall correlations between MODIS AOD over 3 63 pixel groups and ground-level PM2.5 concentrations.

PM2.5

average

AOD

wavelength (mm)

Terra Aqua

r n r n

Hourly 0.47 0.614 3765 0.682 3305

0.55 0.578 3766 0.649 3305

0.66 0.530 3765 0.604 3295

Three hour 0.47 0.630 3806 0.685 3332

0.55 0.593 3807 0.653 3332

0.66 0.543 3806 0.607 3322

Daily 0.47 0.616 3822 0.626 3352

0.55 0.580 3823 0.594 3352

0.66 0.531 3822 0.552 3342

Note: All r values are significant at p , 0.001.

Figure 6. Correlation strengths of the MODIS AOD data at

0.47 mm and the ground-level (GL) PM2.5 concentration averages

over different lengths of time (up to 24 h).

Vol. 36, No. 2, April/avril 2010

Pagination not final/Pagination non finale

8 E 2010 CASI

m10-033.3d 27/8/10 19:13:02

Proof/Epreuve

Figure 7. Variation of the correlation between Terra MODIS AOD (a) and Aqua MODIS AOD (b) at the three wavelengths of 0.47, 0.55,

and 0.66 mm with hourly ground-level PM2.5 concentration, and the MODIS AOD data frequency (c) over various months in 2004.

Figure 8. Comparison of the yearly ground-level PM2.5 concentration average map (a) interpolated using the inverse distance weighting

(IDW) method and the yearly Terra MODIS AOD at 0.47 mm (dimensionless) map (b) interpolated from the grid points.

Canadian Journal of Remote Sensing / Journal canadien de teledetection

Pagination not final/Pagination non finale

E 2010 CASI 9

The underlying mechanism should be addressed in future

research to obtain a better understanding of the relationship

between AOD and PM2.5. The impact of particle size on the

MODIS AOD and PM correlation is also worth exploring in

the future. Locations with heavy aerosol loadings suggestedby MODIS AOD data are likely to have higher ground-level

PM2.5 concentrations and therefore deserve more attention

from the environment protection agencies.

References

Al-Saadi, J., Szykman, J., Pierce, R.B., Kittaka, C., Neil, D., Chu, D.A.,

Remer, L., Gumley, L., Prins, E., Weinstock, L., MacDonald, C.,

Wayland, R., Dimmick, F., and Fishman, J. 2005. Improving national

air quality forecasts with satellite aerosol observations. Bulletin of the

American Meteorological Society, Vol. 86, pp. 1249–1261.

Arya, S.P. 1998. Air pollution meteorology and dispersion. Oxford University

Press, New York.

Barnes, W.L., Pagano, T.S., and Salomonson, V.V. 1998. Prelaunch char-

acteristics of the Moderate Resolution Imaging Spectroradiometer

(MODIS) on EOS-AM1. IEEE Transactions on Geoscience and Remote

Sensing, Vol. 36, No. 4, pp. 1088–1100.

Bone, R.M. 2005. The regional geography of Canada. Oxford University

Press, Don Mills, Ont.

Chu, D.A., Kaufman, Y.J., Zibordi, G., Chern, J.D., Mao, J., Li, C., and

Holben, B.N. 2003. Global monitoring of air pollution over land from

EOS-Terra MODIS. Journal of Geophysical Research, Vol. 108, D21,

No. 4661. doi:10.1029.

Engel-Cox, J.A., Holloman, C.H., Coutant, B.W., and Hoff, R.M. 2004.

Qualitative and quantitative evaluation of MODIS satellite sensor data

for regional and urban scale air quality. Atmospheric Environment,

Vol. 38, No. 16, pp. 2495–2509.

Gupta, P., Christopher, S.A., Wang, J., Gehrig, R., Lee, Y., and Kumar, N.

2006. Satellite remote sensing of particulate matter and air quality assess-

ment over global cities. Atmospheric Environment, Vol. 40, No. 30,

pp. 5880–5892.

Hutchison, K.D. 2003. Applications of MODIS satellite data and products

for monitoring air quality in the state of Texas. Atmospheric Environment,

Vol. 37, No. 17, pp. 2403–2412. doi:10.1016/S1352-2310(03)00128-6.

Hutchison, K.D., Smith, S., and Faruqui, S.J. 2005. Correlating MODIS

aerosol optical thickness data with ground-based PM2.5 observations

across Texas for use in a real-time air quality prediction system. Atmo-

spheric Environment, Vol. 39, No. 37, pp. 7190–7203. doi:10.1016/j.atmo-

senv.2005.08.036.

Hutchison, K.D., Faruqui, S.J., and Smith, S. 2008. Improving correlations

between MODIS aerosol optical thickness and ground-based PM2.5

observations through 3D spatial analyses. Atmospheric Environment,

Vol. 42, No. 3, pp. 530–543. doi:10.1016/j.atmosenv.2007.09.050.

Ichoku, C., Kaufman, Y.J., Remer, L.A., and Levy, R. 2004. Global aerosol

remote sensing from MODIS. Advances in Space Research, Vol. 34, No. 4,

pp. 820–827. doi:10.1016/j.asr.2003.07.071.

Kappos, A.D., Bruckmann, P., Eikmann, T., Englert, N., Heinrich, U.,

Hoppe, P., Koch, E., Krause, G.H.M., Kreyling, W.G., and Rauchfuss,

K. 2004. Health effects of particles in ambient air. International Journal of

Hygiene and Environmental Health, Vol. 207, No. 4, pp. 399–407. doi:10.

1078/1438-4639-00306 PMID:15471105.

Kaufman, Y.J., Tanre, D., Remer, L.A., Vermote, E.F., Chu, A., and Hol-

ben, B.N. 1997. Operational remote sensing of tropospheric aerosol over

land from EOS moderate resolution imaging spectroradiometer. Journal

of Geophysical Research, Vol. 102, D14, pp. 17 051–17 067. doi:10.1029/

96JD03988.

Kumar, N., Chu, A., and Foster, A. 2007. An empirical relationship between

PM2.5 and aerosol optical depth in Delhi Metropolitan. Atmospheric

Environment, Vol. 41, No. 21, pp. 4492–4503. doi:10.1016/j.atmosenv.

2007.01.046.,

Levy, R.C., Remer, L.A., Mattoo, S., Vermote, E.F., and Kaufman, Y.J.

2007. Second-generation operational algorithm: retrieval of aerosol

properties over land from inversion of Moderate Resolution Imaging

Spectroradiometer spectral reflectance. Journal of Geophysical Research,

Vol. 112, D13211. doi:10.1029/2006JD007811.

Li, H., Faruque, F., Williams, W., Al-Hamdan, M., Luvall, J., Crosson, W.,

Rickman, D., and Limaye, A. 2009. Optimal temporal scale for the cor-

relation of AOD and ground measurements of PM2.5 in a real-time air

quality estimation system. Atmospheric Environment, Vol. 43, No. 28,

pp. 4303–4310. doi:10.1016/j.atmosenv.2009.06.004.

Masuoka, E., Fleig, A., Wolfe, R.E., and Patt, F. 1998. Key characteristics

of MODIS data products. IEEE Transactions on Geoscience and Remote

Sensing, Vol. 36, No. 4, pp. 1313–1323. doi:10.1109/36.701081.

Neuberger, M., Schimek, M.G., Horak, F., Moshammer, H., Kundi, M.,

Frischer, T., Gomiscek, B., Puxbaum, H., and Hauck, H. 2004. Acute

effects of particulate matter on respiratory diseases, symptoms and func-

tions: epidemiological results of the Austrian Project on Health Effects of

Particulate Matter (AUPHEP). Atmospheric Environment, Vol. 38,

No. 24, pp. 3971–3981. doi:10.1016/j.atmosenv.2003.12.044.

Pope, C.A., III, Verrier, R.L., Lovett, E.G., Larson, A.C., Raizenne, M.E.,

Kanner, R.E., Schwartz, J., Villegas, G.M., Gold, D.R., and Dockery,

D.W. 1999. Heart rate variability associated with particulate air pol-

lution. American Heart Journal, Vol. 138, No. 5, pp. 890–899. doi:10.

1016/S0002-8703(99)70014-1 PMID:10539820.

Raj, P.E., Saha, S.K., Sonbawne, S.M., Deshpande, S.M., Devara, P.C.S.,

Rao, Y.J., Dani, K.K., and Pandithurai, G. 2008. Lidar observation of

aerosol stratification in the lower troposphere over Pune during pre-

monsoon season of 2006. Journal of Earth System Science, Vol. 117,

No. S1, pp. 293–302.

Remer, L.A., Tanre, D., and Kaufman, Y. 2006. Algorithm for remote sens-

ing of tropospheric aerosol from MODIS: collection 5. MODIS Algorithm

Theoretical Basis Document. Available at http://modis-atmos.gsfc.nasa.

gov/MOD04_L2/ atbd.html.

Seinfeld, J.H. 1986. Atmospheric chemistry and physics of air pollution.

Wiley, New York.

Tian, J., and Chen, D. 2007. Evaluating satellite-based measurements for

mapping air quality in Ontario, Canada. Journal of Environmental Infor-

matics, Vol. 10, No. 1, pp. 30–36. doi:10.3808/jei.200700097.

USEPA. 2003. Air quality index — a guide to air quality and your health. US

Environmental Protection Agency (USEPA), Washington, D.C.

van Donkelaar, A., Martin, R.V., and Park, R.J. 2006. Estimating ground-

level PM2.5 using aerosol optical depth determined from satellite remote

sensing. Journal of Geophysical Research, Vol. 111, D21201. doi:10.1029/

2005JD006996.

Villeneuve, P.J., Goldberg, M.S., Krewski, D., Burnett, R.T., and Chen, Y.

2002. Fine particulate air pollution and all-cause mortality within the

Harvard six-cities study: variations in risk by period of exposure. Annals

m10-033.3d 27/8/10 19:13:04

Proof/Epreuve

Vol. 36, No. 2, April/avril 2010

Pagination not final/Pagination non finale

10 E 2010 CASI

of Epidemiology, Vol. 12, No. 8, pp. 568–576. doi:10.1016/S1047-

2797(01)00292-7 PMID:12495830.

Wang, J., and Christopher, S.A. 2003. Intercomparison between satellite-

derived aerosol optical thickness and PM2.5 mass: implications for air

quality studies. Geophysical Research Letters, Vol. 30, No. 21, 2095.

doi:10.1029/2003GL018174.

Yap, D., Reid, N., De Brou, G., and Bloxam, R. 2005. Transboundary air

pollution in Ontario. Ontario Ministry of the Environment, Toronto, Ont.

m10-033.3d 27/8/10 19:13:05

Proof/Epreuve

Canadian Journal of Remote Sensing / Journal canadien de teledetection

Pagination not final/Pagination non finale

E 2010 CASI 11