During the last decades, satellite technology has been outstandingly
improved, providing huge amounts of Earth Observation (EO) data to
be processed and stored. The availability of very high resolution
sensors has encouraged the birth of new domains for remote sensing
applications. Relatively new fields in this frame are Image
Information Mining (IIM) and Content Based Image Retrieval (CBIR).
These fields are born to provide solutions for querying very large
EO archives by content. This dissertation tries to contribute on the
IIM domain, providing new image processing algorithms and
optimization processes for mining image databases.
The study of an IIM system can be focused on signal processing
methods, data compression, semantic knowledge discovery,
human-machine interaction or system architecture design. Thus, the
system can be divided in three modules: on one hand, we have the
off-line tasks, consisting of signal and image processing methods.
The extracted information of these algorithms is based on a
hierarchical Bayesian representation, and usually, is very time
consuming. On the other hand, we explore the on-line actions that
are performed at real time through an interaction with the user.
Finally, an optimal software architecture where all these concepts
are merged has to be studied. In this thesis, contributions on these
three modules are provided.
We begin studying multi temporal high resolution image analysis
under different illumination conditions and strong background
clutter. The aim is to build a target detection map through a
synergy of image processing methods. However, we can be faced with a
common problem while extracting information from EO data, which is
the estimation of parameters. Often the accuracy of the methods is
strongly dependent on the selection of parameters and it is
difficult to a priori know the optimum one. This is the motivation
for the second contribution that deals with this problem. To cope
with it, we implement an algorithm based on clustering features that
uses information and rate distortion theories to help in the
assessment of parameters.
One of the main characteristics of an IIM system, is its potential
to learn though human interaction. The user provides some examples
of his interests, and based on them, the system learns his
preferences, searches for them in large archives, and returns
similar contents to the user provided ones. In this framework, we
developed a multiple classifier, that enables the user to provide
more than one example type. Thus, the system will be queried for
different features, refining the query results and search accuracy.
In order to be an operable and useful system, all new features
proposed in this dissertation have to be accomplished in a modular
system architecture. The system, from the software design point of
view, must be opened, standard compliant and accessible though
Internet. The software architecture design of the IIM system is the
last contribution of this thesis. For building the system, we have
to consider the following aspects:
- how to manage the large data volume of original and processed
images;
- the automatization of tasks as loading new data, extracting features
or generation of thematic maps;
- how to adapt the system to the user knowledge, that is, the image
interpretation has to be adapted to the symbols the users are able
to recognize and to the specific semantics of their domains;
- how to perform the man-machine communication through a continuous
interaction and exchange of knowledge.