VERSEAS TRAVEL REPORT
IGARSS97 - Singapore - 4-8 August 1997
N A Campbell
There were five main areas of interest
to me:
1. Data fusion/neural nets
2. Bi-directional reflectance (BRDF) modelling
3. Image processing
4. Radar applications, especially environmental
5. The commercial exhibition.
1. Data fusion/neural nets
I found the papers in these areas somewhat
disappointing, largely (with hindsight) because the Project has
active research interests in these areas.
The efforts in data fusion were largely in two areas: combining
images by merging them using variations of the IHS transform (without
realising that they were only forming linear combinations of the
input bands); and neural nets, to bypass the distributional problems
that arise with data from different sources.
While the use of neural nets may be expedient, it is possible
to combine data sources using statistical approaches; conditional
probability networks offer one possibility.
In one paper, radial basis function neural nets were used to integrate
radar data sources; the author did not acknowledge that the approach
involved fitting a mixture of Gaussians with (common) diagonal
covariance matrix. The main problem associated with fitting multivariate
mixtures of Gaussians - where to locate the initial nodes (and
how many) - still exists when using radial basis functions. And
mixture models are sensitive to atypical observations.
The most interesting paper in the area was by David Goodenough and colleagues who used a system of experts for intelligent data management. The work was more computer-oriented than mathematical, and the results were impressive.
2. Bi-directional Reflectance (BRDF)
There is considerable interest within
the CSIRO EOC on BRDF modelling. This arises naturally in two
areas: (i) seamless mosaicing of airborne video and aerial photo
images; and (ii) producing standardised, consistent sequences
of AVHRR imagery.
Our interest is in the first area, and very much at the applied
end; we are collaborating with David Jupp by considering statistical
aspects of fitting various (combinations of) kernel models to
airborne video data.
There were several talks and posters relating to BRDF, including
two modelling approaches which could be useful. The first was
by John Dymond and David Pairman and colleagues at Landcare Research
in New Zealand, while the second was by Chen and Leblanc from
CCRS.
There were also papers/posters from the University of Maryland
on temporal BRDF (Liang et al; Kalluri; Rahman et al) and University
of Arizona on compositing AVHRR (W van Leeuwen).
3. Image Processing
A paper (and related poster) on combining
data at different scales used an unmixing approach and gave a
very useful reference to a paper by Forrest Hall and colleagues
that I wasn't aware of previously.
A very useful overview of trends in information processing for
remote sensing was given by C H Chen. He is the editor of IEEE
Transactions on Geoscience and Remote Sensing. The review covered
in part what he considered were the most important areas and papers
to appear in that journal; included were mixture models, robust
statistics and contextual classification.
4. Radar
My initial interest in attending IGARSS97
was because of the sessions on data fusion and neural nets. There
are generally about 10 parallel sessions, so scheduling is a challenge.
I attended a few sessions on radar, mainly to get a feel for the
area. I also plan to read through the proceedings. Our developing
interest is in the potential use of radar for environmental monitoring.
An ERS publication at ISPRS in Vienna showed a promising application
of multi-temporal ERS-SAR data to identify different cover types.
E Nezry presented a paper at IGARSS97 on the early identification
of the area sown to agricultural crops using ERS SAR data in which
a series of images relatively early in the growth cycle gave promising
predictions. There seems considerable scope for more effective
use of the time-sequence data.
Considerable attention is being paid to the segmentation of SAR
images. Several papers involved the use of Markov random fields
(MRFs) (remember: I'm a novice when it comes to SAR data!), including
a very nice presentation by Y Dong of the University of New South
Wales involving wavelets and MRFs for thresholding and segmentation,
and one by PC Smits and S G Dellepiane involving a gamma model
for the SAR data and a MRF regularisation term.
5. Commercial Exhibition
The commercial exhibition included two
relating to the coming generation of high spatial resolution satellites,
and several relating to image processing systems.
John Douglas, the Australian representative for Earth Watch (Early
Bird, Quick Bird), was at the Conference.
One of the highlights of the Conference for me was the time I
spent with Philip Cheng, the Senior Software Engineer with PCI.
Some recent work in Perth with Changming Sun involved developing
robust procedures for calculating relative and absolute orientation
as part of the procedure for determining DEMs from overlapping
aerial photographs. The proposed SW Land Monitor project will
include extensive generation of DTMs, and part of the NHT proposal
includes acquisition of suitable soft-copy software by DOLA for
quality control.
The PCI modules relating to the soft-copy calculations were very
easy to use. I had two test images on a CD; these were loaded
onto the PCI system and a DTM was generated. I also used the results
from the PCI system to cross-check the software I wrote as part
of the work with Changming.
DOLA and CMIS have been carrying out an evaluation of soft-copy
systems using a test data set generated explicitly for the evaluation.
PCI will carry out similar calculations to those already provided
by a number of vendors.
I also spent considerable time discussing algorithmic aspects
of image registration and image matching, including the use of
collinearity equations and orbital geometry for image rectification.
This is important as we move towards producing accurate geometric
mosaics over the SW of WA.