In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
(a) Water(Wa) 
(e) Wo+Cr 
(f) Wa+Wo (g) Wa+Bu+Wo 
Figure 2: Some training samples annotated with keywords. 
(b) Woodland(Wo) 
(c) Building(Bu) 
(d) Cropland(Cr) 
(h) Cr+Wo+Wa 
(e) MMAM-A11 
(f) MMAM-Hist (g) MMAM-GMRF (h) MMAM-Huy 
Woodland Water Building Cropland 
Figure 3: (a) Quantitative evaluation area of 1400 x 1200 pixels; (b) The corresponding hand-labeled ground truth; (c) Classification 
result using pixel-based Wishart ML; (d) Classification result using patch-based Wishart ML (patch size:20 x 20); (e) Classification 
result using MMAM with all features;(f) Classification result using MMAM with histogram features ;(g) Classification result using 
MMAM with GMRF features ;(h) Classification result using MMAM with Huynen decomposition features. 
It has been tested and validated on a large RadarSat-2 PolSAR 
scene image classification task, and produces satisfactory clas 
sification results, it outperforms traditional Wishart ML meth 
ods with detailed pixel-level labeled training data, even when us 
ing only one feature-multichannel histogram. Moreover, we use 
the over-segmentation based soft assignment techniques (Patch to 
Pixel labels mapping) to reduce the block effect in each subimage 
and improve the visual effects. While the results presented here 
are encouraging, there is still a need for further improvements. 
Future extensions would be the introduction of other sources of 
contextual information like scale information and the combina 
tion with more informative feature descriptors.