Radar images for rainfall measurements: a neural network analysis

Motivated by a real world problem, this study develops a neural network approach to identify and evaluate the relationship between atmospheric radar reflectivity and ground level rainfall intensity. Rainfall is one of the most difficult elements of hydrologic cycle to measure and forecast. This is due to the tremendous range of variability it displays over a wide range of scales both in space and time. Weather radar constitutes an attractive possibility for improving the description of rainfall fields as it can provide high resolution images in space and time of the atmospheric reflectivity over large areas. Radars emit short pulses of energy in the radio-frequency portion of the electromagnetic spectrum, which are focused by the antenna into a narrow beam. From the backscattering energy of the hydrometeors that returns to the transmitter it is possible to obtain estimates of the rainfall field. Radar data are displayed on constant altitude plan position indicators (CAPPIs) levels. The empirical Marshall–Palmer (MP) relationship is normally used in operational hydrology to link together the reflectivity factor Z at the lowest CAPPI level and rainfall intensity R at the ground level. The coefficients reflect the dependence of Z from the number and size distribution of meteors present in the volumes scanned by radar beam. The MP relationship needs a careful preprocessing phase to remove known anomalies in the data, which are due to several factors such as, for example, the distribution of water particles, low level precipitation and low level evaporation. In addition, much noise may affect radar data, owing to radar calibration, signal attenuation, and electromagnetic signal propagation path. The preprocessing stage may limit the many applications which require a real time estimation of the rainfall field. Furthermore, the MP relationship exploits the radar image at the lowest CAPPI level, while the importance of incorporating the entire vertical profile of Z to improve the estimates of R has been recognised in several works.These considerations raise the following statistical problem: perform a fast multivariate analysis of the Z-R relationship with a view to making real time predictions and filtering the effect of noise and bias and the presence of outliers in the data. The aim of the neural network analysis performed in this work is to stress the differences between the univariate and the multivariate approach to the Z-R relationship and to compare neural networks with the classical MP relationship, in the univariate analysis, and with other flexible non linear statistical methods, in the multivariate analyisis. Performances are evaluated by means of an empirical study.