In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
427 
time-series of uncertainty values over a year. It has been 
observed that the shapes of uncertainty time-series closely 
match with that of AOD time-series because the uncertainty 
values are highly correlated with AOD. So, we used uncertainty 
estimates as a proxy for the actual AOD labels that cannot be 
observed over candidate sites. A potential drawback of the 
previous three site selection algorithms is that selection by one 
metric is not guaranteed to be the same as that selected by 
another metric. Therefore, we modified these algorithms to 
combine uncertainty, spatial and temporal correlation criteria in 
a single measure. Our objective was to evaluate which approach 
is the most appropriate for AERONET site selection 
2.5 Discovering Correctable AOD Retrieval Error 
We analyzed performance of the operational MODIS aerosol 
retrieval algorithm. Overall, the main sources of MODIS 
aerosol retrieval errors are the separation of surface and 
atmospheric components of the observed radiances, the 
inaccuracies in the forward-simulation model, and inversion 
errors. Some sources of retrieval uncertainties, such as bright 
surfaces or cloud-contaminated scenes, are due to the 
limitations of the MODIS instrument and cannot be corrected, 
while others, such as imperfections in the retrieval algorithm, 
are correctable. Aerosol scientists’ major goal is to understand 
the primary sources of correctable retrieval errors and to use 
such knowledge to improve the retrieval algorithms. The goal of 
this study was to explore if data mining could facilitate this 
process. 
Our approach consisted of the three main components: 1) use 
collocated AERONET and MODIS data to train neural 
networks for the retrieval of AOD; 2) compare the accuracy of 
neural networks and the MODIS operational algorithm, and 3) 
understand the present conditions in instances when the neural 
network is more accurate than MODIS retrievals. A neural 
network trained in the first step is a completely data-driven 
retrieval algorithm, distinct from the model-driven MODIS 
operational algorithm. The drawback of neural network retrieval 
is that its high accuracy is not guaranteed for the conditions 
unlike those at the AERONET sites. As such, neural networks 
are not a completely viable alternative to model-driven retrieval 
algorithms. However, if neural networks can achieve higher 
retrieval accuracy over the AERONET locations, then it is clear 
that the accuracy of a model-driven algorithm can be further 
improved. 
2.6 Unifying Multiple Retrievals by Structural Regression 
The aerosol data are characterized by strong spatial and 
temporal dependencies. To exploit these dependencies we have 
recently developed Continuous Conditional Random Fields 
(CRF) for AOD retrieval that are able to exploit by defining 
interactions among outputs using feature functions 
(Radosavljevic et al, 2010b). The use of features to define the 
CRF models allowed us also to include arbitrary properties of 
input-output pairs into the compatibility measure. Our CRF 
probabilistic model for structured regression uses multiple non- 
structured predictors as its features. Features were constructed 
as squared prediction errors of deterministic and statistical 
models and we showed that this results in multivariate Gaussian 
conditional P(y|x) distribution. Consequently, in the proposed 
approach learning is a convex optimization problem with a 
global solution for a set of parameters and inference is 
conveniently conducted through matrix computation. 
3. RESULTS 
3.1 Spatio-Temporal Data Partitioning for AOD Retrieval 
Following methodology summarize in Section 2.1 we 
performed large scale experiment using 2 years of data from 
more than 200 ground based AERONET sites located at six 
continents spatio-temporally collected with data from MODIS 
instrument aboard NASA’s Earth observing Terra and Aqua 
satellites. The obtained soft partitioning results (illustrated at 
Fig. 4) were compared to the data partitioning used in the 
MODIS operational algorithm that divides the world into three 
spatial-temporal regions based on domain knowledge. The 
experiments showed that the new soft partitioning of Earth 
results in significant AOD retrieval accuracy improvements 
(Radosavljevic et al, 2008). 
01-Jan-2006 
-180 -120 -60 0 60 120 180 
Figure 4: Spatio-temporal partitioning of Earth discovered by 
competition of two AOD prediction models. Pixel color 
corresponds to weight w assigned to one AOD predictor in a 
mixture. The other predictor has weight 1-w. Top panel: winter 
partitions, Bottom panel: summer partitions. 
3.2 AOD Retrieval across Multiple Accuracy Measures 
Neural networks from the ensemble described in Section 2.2 
were trained using collocated data points whose attributes were 
derived from MODIS instrument satellite observations and 
whose target AOD variable was obtained from the ground-based 
AERONET instruments. Instead of relying on MSE 
minimization criterion for neural network training, we used the 
relative error REL, which can be considered as generalization of 
MSE. 
We observed that REL criterion allowed us to achieve increased 
accuracy over certain ranges of AOD values. To provide a 
predictor that is accurate over the whole range of AOD values 
for each of the 5 commonly used accuracy measures, we 
developed an ensemble of neural networks with adaptive cost 
functions. Some networks in the ensemble were specialized in 
predicting small AOD while others were specialized in 
predicting large AOD. The experiments showed that the 
proposed ensemble outperformed an ensemble that used 
standard MSE optimization; it managed to achieve as high 
MSE, R 2 and CORR accuracies while it significantly improved 
MSRE and FRAC accuracies. In addition, AOD prediction 
accuracy of the proposed ensemble was compared to the