The recent increase in Earth Observation (EO) missions has resulted in unprecedented volumes of multi-modal data to be processed, understood, used and stored in archives. The advanced capabilities of satellite sensors become useful only when translated into accurate, focused information, ready to be used by decision makers from various fields. Two key problems emerge when trying to bridge the gap between research, science and multi-user platforms: (1) The current systems for data access permit only queries by geographic location, time of acquisition, type of sensor, but this information is often less important than the latent, conceptual content of the scenes; (2) simultaneously, many new applications relying on EO data require the knowledge of complex image processing and computer vision methods for understanding and extracting information from the data.
This dissertation designs two important concept modules of a theoretical image information mining (IIM) system for EO: semantic knowledge discovery in large databases and data visualization techniques. These modules allow users to discover and extract relevant conceptual information directly from satellite images and generate an optimum visualization for this information.
The first contribution of this dissertation brings a theoretical solution that bridges the gap and discovers the semantic rules between the output of state-of-the-art classification algorithms and the semantic, human-defined, manually-applied terminology of cartographic data. The set of rules explain in latent, linguistic concepts the contents of satellite images and link the low-level machine language to the high-level human understanding.
The second contribution of this dissertation is an adaptive visualization methodology used to assist the image analyst in understanding the satellite image through optimum representations and to offer cognitive support in discovering relevant information in the scenes. It is an interactive technique applied to discover the optimum combination of three spectral features of a multi-band satellite image that enhance visualization of learned targets and phenomena of interest. The visual mining module is essential for an IIM system because all EO-based applications involve several steps of visual inspection and the final decision about the information derived from satellite data is always made by a human operator. To ensure maximum correlation between the requirements of the analyst and the possibilities of the computer, the visualization tool models the human visual system and secures that a change in the image space is equivalent to a change in the perception space of the operator. This thesis presents novel concepts and methods that help users access and discover latent information in archives and visualize satellite scenes in an interactive, human-centered and information-driven workflow.