Structured neural network approach for measuring raindrop sizes and velocities

The paper describes a structured neural network solution to a signal processing problem in the meteorological and telecommunications domains. Optical disdrometers measure raindrop sizes and velocities by registering changes in photodiode current as the droplets pass through a collimated light beam. In an improved dual-beam device developed at CETP, feature extraction multilayer perceptrons applied to 20-sample windows of photodiode current provide input to a higher-level network which reconstructs droplet velocities and diameters in real time. In the tests on simulated data, the measurement precision is quite good for droplets as small as .05 mm radius. The algorithm can be executed either directly on the acquisition PC, or on a neural net coprocessor for additional speed-up.