Detection of anomalous propagation echoes in weather radar data using neural networks

We investigate a neural network-based methodology for detection of the anomalous propagation (AP) radar echo. The methodology is devised to cope with the situations when only single scan data are available. The output of the procedure is quantified in four classes corresponding to the upper limits of 25, 50, 75, and 100% of AP echo per scan, The high dimension of the input data space is reduced by feature extraction based on physical considerations. Fractal based, statistical, and wavelet analyses are performed, and their characteristics are used as features. A feedforward neural network is used for classification in the four classes, with a fuzzy strategy used in the network training. We test the methodology on real data and make a comprehensive assessment of the procedure's accuracy based on cross validation.