dc.description.abstract |
Political, environmental, and commercial needs for information on the Earths surface
and
atmosphere drive the development of improved satellite data products. At
visible and thermal
wavelengths the quality of these products is dependent on our
ability to distinguish between clouds and the underlying surface. Unlike oceans,
land surfaces are
highly heterogeneous, containing a wide range of materials, some
of which exhibit similar
spectral properties to cloud, and hence it is much harder to
distinguish between the two. ~
This research
project, supported by the Along-Track Scanning Radiometer (ATSR)
science team at the Rutherford
Appleton Laboratory (RAL), addresses the need
for
improved cloud detection over land surfaces through the development of an unsupervised
cloud detection system for global ATSR-2 scenes over land surfaces.
The thesis details the
development of the first successful un-supervised near-global
cloud detection scheme for ATSR-2 scenes over land surfaces. The scheme developed
operates on ATSR-2 data using a fuzzy set methodology. The level of membership of
the
fuzzy sets is determined using aggregated Gaussian distribution functions defined
in a
knowledge base that has been developed from the International Satellite Cloud
Climatology Project (ISCCP) data sets.
This is the first cloud detection
algorithm that is uniquely customisable to its end
users needs.
Specifically, this is achieved through the use of fuzzy set theory and set
membership grades. This elegant solution to the problem achieves cloud detection
as
oppose to cloud clearing, and its final output retains all the information computed
on
possible classifications of image pixels, thus providing the end user with a true
representation of the imprecision inherent in the real-world data.
A
comprehensive quantitative evaluation and inter-comparison of cloud clearing
schemes is
presented. This showed that with respect to other algorithms (in literature
and
currently under development at RAL) F-CLOUD is one of the frontrunners
in a new
generation of cloud detection algorithms over land surfaces.
The scheme is
highly accurate and has immediate potential applications within the
development programme of future ATSR-2/AATSR products at RAL. Using confusion
matricies to
analyse hardened results yielded a mean classification accuracy of
92.3%
(for a total of forty-five scenes analysed against neph-analysis derived cloud
masks). |
en_UK |