Chair/s: J.S. Lee/ S. Quegan
All papers in this session dealt with classification and segmentation
techniques using polarimetric SAR data.
· A common data set used in several papers is the AIRSAR P-,
L-, C-Band polarimetric SAR data of Flevoland, The Netherlands, but the
EMISAR multitemporal C and L-band dataset from Denmark is a very important
· A new classification scheme (Dr. Hoekman), that transforms
the covariance matrix into 9 intensity components, produces results with
high classification rate as claimed by the authors. This algorithm
uses segmented data and a model for class statistics (not just the Wishart
distribution, which is suitable for field statistics). The results
were obtained on the particular Flevoland test site and further verification
of the high classification rate and robustness is required on other test
· Three principal classification approaches were used: hierarchical
decision trees (T. Le Toan, S. Quegan and H. Skriver), feature-based ML
and the Wishart classifier (J.S. Lee, E. Pottier and L. Ferro-Famil), the
latter two perhaps linked to an ISODATA algorithm (as done in the paper
by Quegan et al.). The Flevoland data indicated that both techniques
have their merits, depending on the crop conditions and hence period in
the growing season.
· Pixel-based, (GIS) area-based and (automatic) segmentation-based
methods were used. Pixel-based methods give good results when state-of-the-art
filters are used. The best results used area-based methods; a number of
methods are available to segment polarimetric images.
· Partial polarimetric data (dual polarizations) can be used
to separate winter from spring crops, although the best separation used
the HH-VV correlation coefficient.
· A POLSAR analysis and classification software package: POLSAR
PRO, contains very up-to-date algorithms. This package is developed
by University of Rennes under the supervision of E. Pottier. This
software package is free.
· The paper by X. Blaes was cancelled.
1. Is fully polarimetric data necessary? Which frequency is the
best for land-agriculture applications?
Yes, fully polarimetric data is necessary for land-agriculture applications.
There was some discussion about the relative merits of L-band (Lee’s view)
or C-band (Quegan et al.’s opinion). However, L-band is preferred
also for other applications such as sea ice and forest (Lee’s opinion).
2. How to verify the high classification rate of Dr. Hoekman’s classification
algorithm and other algorithms?
Recommend establishing common data sets for evaluating classification
algorithms. Their purpose is to test robustness (transferability) of algorithms
and to compare algorithms.
· The Flevoland data used by many papers in this session has
some merits for this purpose, but has rather special properties and cannot
be considered typical even of European agriculture. Data sets from other
sites are needed.
· Multi-temporal data are needed to evaluate the classification
capabilities at various crop growth stages.
· Training sets together with well-defined and trustworthy ground
data need to be pre-defined for classification evaluation and for fair
comparisons of algorithms (Quegan’s comment: this depends on the training
strategy, which might be algorithm specific).
· The rules for comparison need to be clearly stated, e.g.,
no tweaking of parameters, full processing chain specified and transferable,
training and learning process fully specified, etc...
· Computational efficiency is another important comparison for
As well as providing procedures for evaluating and comparing algorithms,
effort must go into understanding WHY the best algorithms are the best.
3. Besides crop classification, what can additional crop parameters
(height, biomass, plant structure, etc.) be reliably extracted using POLSAR
and Pol-INSAR data?
· POLSAR can be used winter crop and spring crop separation,
and broad leaf (random) versus cereal (structured) crop types. There
is some evidence that partially polarimetric SAR can be estimate biomass
through the HV/HH and/or VV/HH ratios; this needs to be tested with ASAR.
· It may be possible to extract crop height using POL-INSAR
technique. More research needs to be done in this area, and C-band
is probably a better frequency to use than L-band.
4. Does POLSAR supervised classification performs better than unsupervised
classification? What are their merits?
· In general, supervised classification has better classification
rate than unsupervised classification.
· Positive crop identification.
· Classes of close scattering properties may be separated, but
they could be grouped into one class in the unsupervised classification.
· Disadvantage: Needs ground truth collection to establish training
· Automated techniques easy to apply. No ground truth
collection is needed.
· Scattering characteristics (even bounce, odd bounce, volume
scattering, entropy, anisotropy, etc) of each class may provide hints for
· Distinct classes not corresponding to predetermined crop types
· Disadvantage: No positive crop identification. Classes
separated using supervised techniques may be grouped into one.
5. What can ENVISAT ASAR with C-band dual polarizations achieve in
· Biomass (HV/HH and VV/HH).
· Rice crop monitoring.
· Separation of cereal versus root crop (HH/VV).
with multitemporal polarimetric SAR data
Professor Shaun Quegan