or
ts
nd
except several most difficult classes (these classes are usually
hard to label and are mainly confused, see Tables 3 and 4). The
classes 7 (Grapes_untrained) and 14 (Vinyard_untrained) illus-
trate the highest confusion with classification accuracies equal
to 68.79% and 65.05% (Table 4), respectively. Classes 11 (Let-
tuce romaine. 5wk) and 8 (Soil. vinyard develop), also class 13
(Lettuce romaine. 7wk) and 12 (Lettuce-romaine-6wk) and 11 (Let-
tuce romaine. 5wk) are less confused.
Median filtering reduced the influence of the outlier samples in
the input data for classification. Confusion among classes was re-
duced and a better labeling was reached. MNF data preprocessing
allows to reduce the time of calculation with a competitive clas-
sification accuracy. On full bands set data a better classification
accuracy is expected to be obtained.
Approximate inference methods should be employed for the like-
lihood probability computation. Approximate inference allows to
calculate decisions with the accuracy comparable to the results of
full propagation methods but with a high reduction of run time.
In this work Mean Field approximate inference method was em-
ployed. Factor graph allows to perform an inference for one class
(to produce a probability map) leading to an application of mate-
rial detection in hyperspectral data.
Among disadvantages we can note that probabilistic models re-
quire computational time higher than many classification meth-
ods, since inference in each point of input data is performed.
Also maximum principle on the likelihood probability maps (per-
formed to obtain class label map) can be a source of misclassifi-
cation.
Table 2: The accuracy of salinas benchmark classification using
the FG (MNF 20 features, alphabet size: 100). Additional ex-
periment with feature median filtering is also presented. OVA —
overall accuracy, Kappa — Cohen's Kappa
Method OVA, % | Kappa
FG 81.3692 | 0.7921
FG (Median 5 x 5) | 85.3217 | 0.8358
(a) (b) (c)
Figure 3: Classification maps for Salinas benchmark using factor
graphs: (a) MNF, 20 features, alphabet size: 100, (b) MNF, 20
features, median filtering 5 x 5, alphabet size: 100, (c) ground
truth label map
4 CONCLUSIONS
The paper presents another successful area of factor graphs ap-
plication: multispectral data classification. A relatively simple
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
structure of the FG allows to reach a competitive accuracy of
classification even on data with decreased radiometric range (rep-
resented on the alphabet). An important property of factor graph
classification is that the method requires a relatively low number
of training samples (only 20 points for a class). Separate process-
ing of input features (spectral bands) and employment of the pre-
sented data fusion and classification model is not influenced by
the limitations of data dimensionality (i.e. there is no the curse of
dimensionality). Classification on full data (all spectral bands) is
possible to run (comparing to MNF features) and will take more
computational time.
REFERENCES
Bishop, C., 2006. Pattern recognition and machine learning. In-
formation Science and Statistics. Springer, New York.
Boardman, J. and Kruse, F., 194. Automated spectral analysis: a
geological example using aviris data, north grapevine mountains,
nevada. In: Proceedings of the Tenth Thematic Conference on
Geological Remote Sensing, San Antonio, TX, USA, pp. 407—
418.
Bratasanu, D., Nedelcu, I. and Datcu, M., 2011. Bridging the se-
mantic gap for satellite image annotation and automatic mapping
applications. IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing 4(1), pp. 193—204.
Frey, B. and Jojic, N., 2005. A comparison of algorithms for
inference and learning in probabilistic graphical models. IEEE
Transactions on Pattern Analysis and Machine Intelligence 27(9),
pp. 1392-1416.
Green, A., Berman, M., Switzer, P. and Craig, M., 1988. A trans-
formation for ordering multispectral data in terms of image qual-
ity with implications for noise removal. IEEE Transactions on
Geoscience and Remote Sensing 26(1), pp. 65-74.
Kschischang, F., Frey, B. and Loeliger, H.-A., 2001. Factor
graphs and the sum-product algorithm. IEEE Transactions on
Information Theory 47(2), pp. 498—519.
Lienou, M., Maitre, H. and Datcu, M., 2010. Semantic annota-
tion of satellite images using latent Dirichlet allocation. IEEE
Geoscience and Remote Sensing Letters 7(1), pp. 28-32.
Wang, C., Blei, D. and Fei-Fei, L., 2009. Simultaneous image
classification and annotation. In: IEEE Conference on Computer
Vision and Pattern Recognition, pp. 1903-1910.
