In: Wagner W., Szikely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
425 
The main limitation of ground based observations is their low 
spatial coverage. Therefore, satellite-based aerosol related 
observations from several new instruments of high spatial 
coverage are also considered. The MODerate resolution 
Imaging Spectrometer (MODIS), aboard NASA’s Terra and 
Aqua satellites launched in 1999 and 2002 with a single camera 
observes reflected solar radiation from the Earth over a large 
spectral range in 7 bands. It has a repeating cycle of 16 days and 
high spatial resolution with almost daily coverage of the entire 
planet. In comparison, MISR, also aboard Terra satellite, is a 
nine camera instrument with four bands per camera that 
provides global coverage every 9 days. MISR collects raw data 
at 1.1 km resolution, but retrieves aerosol properties at 17.6 km 
resolution for twenty-four postulated aerosol types. MODIS on 
the other hand collects data at 1 km resolution and its retrievals 
are provided at 10km resolution. 
Designing accurate AOD predictors from satellite observations 
is a very challenging task due to various problems including 
reflectance superposition from multiple sources (effects of 
clouds and surface reflectance are illustrated at Fig. 2). 
Therefore, satellite based retrievals are less accurate than 
ground based retrievals. However, they provide high spatial 
coverage and so are very important for climate studies. 
Figure 2: Physics of satellite-based retrieval. 
Most operational aerosol retrieval algorithms are constructed as 
inverse operators of high-dimensional non-linear functions 
derived from forward-simulation models according to the 
domain knowledge of aerosol physical properties [35,36,67,68]. 
For example, MISR uses 24 compositional aerosol models in 
the Aerosol Climatology Product. These aerosol models, such 
as mineral dust, biomass burning particles and urban soot, are 
considered to be representative of the types found over the 
globe. They are mixtures of individual component aerosols, 
where each component is defined by a size distribution, particle 
shape, spectral index of refraction and vertical distribution 
within the atmosphere. Up to three components can comprise an 
aerosol mixture, and the fractional optical depths of the 
components making up a given mixture are pre-specified. For 
each component aerosol, the corresponding radiative properties 
are computed using wavelength, illumination, and view 
geometry information. The results are recorded in a look-up 
table. By using a modified linear mixing theory, the radiative 
properties of a mixture are calculated during the retrieval 
process. These simulated data are then compared to actual 
observations for the appropriate scene type (land or ocean). 
According to a set of goodness-of-fit criteria based on the 
domain knowledge, the matched aerosol model in the look-up 
table is used for AOD computation. 
Drawbacks of deterministic retrieval methods include (1) high 
computational cost due to inversion of nonlinear forward 
models; (2) slow development due to manual construction of 
the postulated physical models; (3) suboptimality due to 
difficulties in capturing complex radiance-aerosol relationships 
in all realistic scenarios; and (4) significant retrieval 
inaccuracies that are due both to the instrument limitations and 
imperfections in the retrieval algorithms. 
Our team has demonstrated that more accurate retrieval is 
achievable by a completely data-driven approach using spatio- 
temporally collocated satellite and ground based observations as 
shown at Fig 3 (Han et al, 2005a; 2005b; 2006a; 2006b; Das et 
al, 2008; Obradovic et al, 2006; Xu et al, 2005; Zhuang et al 
2008). This statistical method consists of training a nonlinear 
regression model using the satellite observations as inputs and 
ground based AOD measurements as target. Some of our recent 
related activities and findings are discussed in this article (Das 
et al, 2009; Radosavljevic et al 2008; 2009; 2010a; 2010b; 
Ristovski et al, 2009; Vucetic et al, 2008). 
Figure 3: Spatio-temporal collocation of MODIS and 
AERONET data. A is an AERONET site with AOD retrieved 
within a short time before and after the satellite overpass 
(circle dots). The square regions are MODIS observations in a 
proximity of AERONET site A at the satellite overpass time. 
2. METHODOLOGY 
The overall objective of our study is to facilitate aerosol 
retrieval algorithms development, application and modification 
by developing data mining methodology that utilizes satellite 
data together with ground-based measurements. Our specific 
aims are do determine if data mining can 
• provide accurate statistical AOD retrievals; 
• help discovering the major sources of correctable 
retrieval errors of deterministic retrievals; 
• improve understanding spatio-temporal properties of 
deterministic retrievals. 
An overview of our recently proposed approaches towards 
achieving these aims is provided in this section. 
2.1 Spatio-Temporal Data Partitioning for AOD Retrieval 
In principle, AOD retrievals of high spatial resolution can be 
obtained from satellite observations by training a regression 
model on a dataset that consists of the satellite observations as 
inputs and more accurate ground based AOD measurements as 
output. However, challenges of such supervised learning on 
aerosol data collected over space and time include existence of 
different relationships among observations and AOD over 
various spatio-temporal regions. In such situations an 
appropriate spatial-temporal data partitioning followed by 
building specialized predictors could often achieve higher 
overall prediction accuracy than when learning a single 
predictor on all the data. In practice, such partitions are 
typically decided based on prior knowledge. 
As an alternative to the domain-based partitioning, we have 
proposed a method that automatically discovers a soft spatio-