Tuesday, 11 January 2000
Kwo-Sen Kuo, Univ. of Alabama, Huntsville, AL; and T. A. Berendes, D. A. Berendes, and R. M. Welch
A new classification scheme using neural networks with back-propagation learning has been developed for classifying meteorological satellite imagery. In our previous investigations, each of the neural network's output nodes corresponds to one and only one of the specific classes predetermined by human experts who select and label the training patterns for the neural network. A test pattern is said to belong to the corresponding class of the output node that yields the maximum response from the trained neural network. Since the imagery in question often contains quite complicated patterns, for example, multi-layered and multi-phased clouds or thin layers of atmospheric particulate matter over complex surfaces, the number of specific classes may get quite large; more than
thirty classes have been used in our previous studies. This not only presents difficulty for the human expert to label the patterns due to the large number of choices, but also makes the neural network less efficient because there has to be as many output nodes as the number of specific classes. In actuality, the majority of the specific classes are composed of a small number of elementary classes.
In this new scheme, we assign a neural network's output node to represent an elementary class with its response signaling the strength of the elementary class's presence in the pattern. The number of output nodes is greatly reduced (from 30+ to 7) and the efficiency of the neural network is consequently enhanced. Preliminary result using training data from the old approach shows an increase of classification accuracy of more than 3%, from 84.3% to 87.5%, using the new method. The performance is expected to increase even more when training patterns are labeled according the configuration of the new approach.
Beside the increased economy of representation and improved neural network efficiency, the new approach also naturally lends itself to the quantitative interpretation of the neural network output signal. Successive refinements
of the training data based on the classification result of the previous iteration will enable the neural network not only to identify the presence of an elementary class but also the percentage of that class in a test pattern.
![](images/flag.png)
- Indicates paper has been withdrawn from meeting
![](images/blue_ribbon.png)
- Indicates an Award Winner