Interactive learning in remote-sensing image databases

In this paper we present a concept of interactive learning in an intelligent remote sensing image database that allows users to query by image content. First, an application-free description of the image content is generated at data insertion. This description should globally capture the structures in the image data. Then, in an interactive step, users of different application fields can define their specific semantic labels based on the application-free representation and use them for later database queries. Such a system might help users of all levels of expertise-both experts and novice users-to find images in the archive that are useful for their particular remote sensing application. We briefly review a hierarchical description of the image content and then sketch the basic step of Bayesian inference used for interactive learning. We present one example of this interactive learning taken from our test database consisting of about 1000 small 1024/spl times/1024 images derived from 11 geocoded Landsat TM scenes.