Intelligent storm identification system using a hierarchical neural network

The paper presents a system to detect severe storm events from non-Doppler radar data using a hierarchical artificial neural network (HANN). The system incorporates three levels of data processing: (i) dimensionality reduction (by data slicing, fragmentation, and preprocessing), (ii) feature extraction in the form of codebooks, and (iii) pattern recognition and classification. We study various schemes of such a processing structure. In one scheme, the first level processing slices the volumetric radar data into a set of images representing radar echo intensity at constant altitudes and fragments the images into image blocks. The second level processing extracts features from the image blocks using a self-organizing feature map (SOFM) neural network. It results in a set of codebooks, which is used by a back propagation (BP) neural network at the third level processing for classification. We have used a limited set of 22 known storm events for our experiments and system development. Preliminary results show 100% correct classification of the storm set. The scheme has been implemented in an X-windows environment, which has been installed in the Atmospheric Environment Services, Winnipeg, Canada, for field tests