Artificial neural network estimation of rainfall intensity from radar observations

Artificial neural networks are used to identify the relationship between weather radar observations of the reflectivity factor Z and rain gauge measurements of rainfall intensity R. These networks are trained and tested using a real data set of reflectivity observed by the Monte Grande weather radar (Teolo, Italy) and rainfall intensity measured by five rain gauges in the Cortina d’Ampezzo area (Italian Dolomites). A principal components analysis is also carried out to investigate the correlation between the Z values at 11 constant altitude plan position indicator levels and to synthesise these values into fewer orthogonal input variables for the networks. Besides the widely used Marshall-Palmer relationship, linear models and flexible discriminants like generalised additive models are used as a benchmark against which the predictive performances of the neural models are measured.Volumetric scans of radar reflectivity Z and gage measurements of rainfall intensity R are used to explore the capabilities of three artificial neural networks to identify and reproduce the functional relationship between Z and R. The three networks are a multilayer perceptron, a Bayesian network, and a radial basis function network. For each of them, numerical experiments are conducted incorporating in the network inputs different descriptions of the space-time variability of Z. Space variability refers to the observations of Z along the vertical atmospheric profile, at 11 constant altitude plan position indicator levels, namely ZT=(Z1,...,Z11). Time variability refers to the observations of Z at the time intervals prior to that for which the estimate of R is provided. Space variability is evaluated by performing a principal component analysis over standardized values of Z, namely Z~, and the first two principal components of Z~ (which describe 91% of the original variance) are used to synthesize the elements of Z into fewer orthogonal inputs for the networks. Network predictions significantly improve when the models are trained with the two principal components of Z~ with respect to the case in which only Z1 is used. Increasing the time horizon further improves the performances of the Bayesian network but is found to worsen the performances of the other two networks.