• About
  • Advanced Search
  • Browse Proceedings
  • Access Policy
  • Sponsor
  • Contact
  • "A NEURAL TECHNIQUE FOR SOLVING CLASSIFICATION PROBLEMS IN REMOTE SENSING"

    Paper ID

    IAF-94-094

    author

    • Tomas Zdravev
    • Hristo Nikolov
    • Doyno Petkov

    company

    Bulgarian Academy of Sciences - Solar-Terrestrial influences Lab.

    country

    Bulgaria

    year

    1994

    abstract

    An object classification problem, where the pixels belongs to an image obtained from a multichannel videospectrometer, in the matter of fact is an object recognition task. In this case, having the spectral irradiance characteristics of the observed by remote sensing techniques objects, is possible to join them to one of the previously formed set of objects with well known and studied properties. The way we do such a reference, always done with preliminary estimated and admitted error, resembles the process of human thinking and perception. The aim of the proposed paper is to formalize the above mentioned process using methods similar to these a man applies about appurtenance of an object to a previously defined set. Structure of the neural network and links in it to form an optimal classification is discussed. Also the achieved degree of compression after classification is conferred. Included is discussion concerning comparison between the implemented by us teaching for the neural structure algorithm with other used for teaching neural nets. Shown are first results for classification of the used for teaching neural net data and data received from calibrated sources. As we consider, such a "human-like" system should be able to perceive and identify more than one object at once, and is possible to implement it in parallel computing environment.