Accession Number : ADA264628


Title :   Neural Network Cloud Classification Research


Corporate Author : MITRE CORP BEDFORD MA


Personal Author(s) : Smotroff, Ira G


Full Text : https://apps.dtic.mil/dtic/tr/fulltext/u2/a264628.pdf


Report Date : Mar 1993


Pagination or Media Count : 69


Abstract : Neural networks are appropriate for meteorological classification tasks for a number of reasons. First, their associative properties allow graceful degradation of performance under conditions of ambiguity and noise, thus avoiding the brittle behavior of many standard approaches. Second, they learn to perform tasks which cannot easily be specified analytically, such as non-linear discriminate functions. Finally, they can be executed in realtime on appropriate hardware. To exploit these properties, this research developed a general approach to meteorological classification based on neural network data fusion. The system was applied to cloud type identification from satellite imagery. The current experiment is one of the first to provide a large cloud database on which to train, and as such is one of the first true cross- validation experiments in this area. While the 27 days of data provides many pixel samples of the cloud types present at a particular hour, the question to be answered here was whether the samples collected on particular types of clouds sufficiently represent the variations of that cloud that can appear on a different day. The promising results point to the applicability of neural networks for automated generation of meteorological products in real time.


Descriptors :   *CLOUDS , *METEOROLOGICAL DATA , DATA BASES , NEURAL NETS , VALIDATION , TIME , IDENTIFICATION , CLASSIFICATION , DATA FUSION , APPROACH , AMBIGUITY , PIXELS , NUMBERS , NOISE , BEHAVIOR , STANDARDS , ARTIFICIAL SATELLITES , VARIATIONS , REAL TIME , NETWORKS , DEGRADATION , FUNCTIONS


Subject Categories : Meteorology


Distribution Statement : APPROVED FOR PUBLIC RELEASE