In: Wagner W„ Szbkely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
recently developed operational MODIS Collection 005 retrieval 
algorithm. Results obtained over the entire globe during the first 
six months of year 2005 showed that the proposed ensemble of 
neural networks was significantly more accurate for all the 
considered accuracy measures (Radosavljevic et al, 2010a). 
3.3 Uncertainty Analysis of AOD Retrieval 
Three years of MODIS data collocated with 201 AERONET 
sites over whole the globe have been used for uncertainty 
analysis of AOD retrieval. The average negative log-predictive 
density (NLPD) of the true targets was used as a measure of the 
quality of uncertainty estimation. Committees of 30 neural 
networks were trained on the subset of sites which have data in 
both training and testing years. The obtained results allowed 
analysis of uncertainty of AOD retrieval at a given site over 
time and also uncertainty comparison at multiple sites. As an 
example, we compared uncertainty of AOD retrieval at Beijing 
site in China vs. Muana Loa site in Hawaii to conclude that 
properties of aerosols are much more stable at Muana Loa than 
in Beijing. By further investigation we found that this discovery 
is consistent with domain experts’ expectations as this site 
serves for calibration statistics of AERONET instruments. 
Our analysis of seasonal uncertainty levels over three years also 
revealed existence of different interesting patterns. For example, 
we compared sites with the highest and the lowest average 
uncertainty over the seasons. The highest uncertainty levels 
occur in Asia over all seasons, in Africa during the winter and 
fall, and in the central part of South America during the 
summer. These levels reach extreme values in summer while for 
other seasons are almost equal. On the other hand, the lowest 
levels of uncertainty appear in North America and Europe 
during winter, summer, and fall, and in South America during 
the spring (Ristovski et al, 2009). 
3.4 Selection of Sites for Ground Based Observations 
Data used in sites reduction experiments were distributed over 
entire globe at 217 AERONET sites during years 2005 and 
2006. We performed training on 2005 data and used 2006 data 
for testing. We considered a scenario when current operational 
AERONET sites have to be reduced by 33% or 66%. In all 
experiments, we started from a set of 30 AERONET sites and 
applied the proposed method and the two alternatives to identify 
a subset of 20 or 10 AERONET sites to be retained. The two 
alternatives included a random selection of sites as well as an 
approach based on spatial distance among the sites was also 
considered. Sites were selected such that their spatial coverage 
was maximized. In our experiments, the proposed accuracy- 
based selection achieved consistently better results than the 
alternatives. Also, accuracy of the proposed site reduction 
method did not change much even after removing 20 of the 30 
AERONET sites. Interestingly, on average, the spatial selection 
strategy performed slightly worse than the random selection 
strategy. According to presented results we concluded that the 
proposed accuracy-based sites reduction method is superior to 
spatially-based and random selection alternatives 
(Radosavljevic et al, 2009). 
Our extensive experiments on globally distributed data over 90 
AERONET sites from the years 2005 and 2006 provide strong 
evidence that sites selected using the algorithm proposed in 
Section 2.5 improve the overall AOD prediction accuracy at a 
faster rate than those selected randomly or based on spatial 
diversity among sites (Das et al, 2009). The evaluation data 
spanned the entire world, ranging from January of 2005 to 
December of 2006. A committee of 20 neural networks was 
used to estimate the model uncertainty, each having 10 hidden 
nodes. Initial training set was created with 700 training points 
from 10 randomly selected AERONET sites. We assumed that 
the AOD values corresponding to the remaining training points 
are unknown. Then we proceeded to select t = (1, 2, ... 20} 
sites in twenty independent experiments using the proposed site 
selection algorithms. The R-squared accuracy was computed on 
the test data before and after the selection of prospective sites. 
Our results show that uncertainty-based site selection gives 
significantly higher accuracy over random and spatial distance- 
based selection (especially when only a few sites are to be 
selected) and marginally higher accuracy than temporal 
distance-based method. Performance improves further if 
correlation among the sites is taken into consideration. A 
comparison between accuracies uncertainty-based and all three 
other methods which selects site based on combination of 
uncertainty, spatial and temporal distance shows that there is 
some improvement in performance due to inclusion of spatial 
and temporal correlation metrics over purely uncertainty-based 
selection. Although the improvement in accuracy is small, it is 
not negligible, keeping in mind the huge variability of AOD 
over the entire earth. 
At the continent scale, we observed that different continents 
favor different selection algorithms. For Europe, random 
selection works quite well because a large number of unlabeled 
sites are from Europe and therefore random selection favors 
sites from Europe. In North America, despite the large number 
of candidate sites, the random selection was not very successful. 
The accuracies for North America generally followed the 
accuracies obtained on overall test set. In South America and 
Africa, spatial selection performed better than other methods. 
Especially in South America, it was able to attain a significantly 
higher accuracy (Das et al, 2009) 
3.5 Discovering Correctable AOD Retrieval Error 
Our approach described in Section 2.5 was applied on 3,646 
collocated MODIS and AERONET observations within the 
continental United States. The results showed that neural 
networks are more accurate than the operational MODIS 
algorithm over the observed locations. A study of differences 
between neural networks and the MODIS algorithm revealed 
interesting findings. For example, NN are more accurate than 
MODIS when the retrieval is contaminated by clouds, snow, or 
water (i.e. NDVI <= 0). These and other discoveries have been 
found to be mostly consistent with expert knowledge and 
revealed some new insights into the MODIS algorithm 
performance (Vucetic et al, 2008). 
3.6 Unifying Multiple Retrievals by Structural Regression 
Experiments were conducted on MODIS data at 50x50 km 2 
resolution collocated with observations at 217 AERONET sites 
during years 2005 and 2006. In nested spatio-temporal 5-cross 
validation experiments we determined parameters 
corresponding to the operational C005 and neural network 
(.NN) retrievals over five continents (Asia and Australia were 
treated as a single region). In our experiments CRF achieved 
significantly better accuracy than either NN or C005 alone. 
In follow-up experiments when using indicator functions for 5 
continents (Asia and Australia were treated as one continent) 
CRF achieved better accuracy than either NN or C005 alone