Automated extraction of metadata from remotely sensed satellite imagery

The Earth Observing System (EOS) is scheduled to be placed in orbit by NASA in the late 1990s. This data must be rapidly archieved and made accessible through a variety of discipline-specific indices. This paper discussed research in the Intelligent Data Management (IDM) project at NASA/Goddard Space Flight Center. Recent improvements to low-level feature detection algorithms in computer science hold promise for performing real-time characterization of images. The IDM group is investigating using neural networks and expert systems to produce a summary of the extent of classes within an image. Problems unique to this research include the lack of ground truth for the training of artificial neural networks, the effects of climatic and seasonal changes in deriving prototypical training sets, the volume of data with respect to its processing and storage, and the varied backgrounds of the system's end-users. We concentrate in this paper on the characterization of images (including multispectral scanner [MSS] and thematic mapper [TM] data) using neural networks and the interpretation of the neural network output by an expert system for subsequent archiving in an object-oriented database.