P16.7
Neural Networks Applications to the Rainfall Rate Extraction in Polarimetric Weather Radar
Gaspare Galati, Tor Vergata University of Rome, Rome, Italy; and C. Baldi and G. Pavan
The paper addresses the problem of the reconstruction of the field of rainfall using polarimetric weather radar. It is well known that at the C band and especially at the X band the reconstruction of the rainfall rate (R) profile along the range using absolute (ZH) and differential (ZDR) reflectivity measurements is significantly affected by attenuation (aH, aD). This problem was extensively studied in the past and iterative attenuation correction techniques based on a cumulative procedure were developed, in which the attenuation at nth cell is estimated using the attenuation corrected reflectivity values at previous (n-1)th cell. The effectiveness of iterative attenuation correction techniques is mainly limited by the propagation of errors along the path. This paper considers two potential uses of neural networks to solve the two limitations of iterative attenuation correction techniques. First, relatively simple neural networks have been used to reconstruct the rainfall rate in a single resolution cell. This way is alternative to the widely-used parametrizations using the absolute and differential reflectivities or the differential phase constant: R=a ZH^b ZDR^c R=a' KDP^b' Second, with more complex neural networks an attempt has been made to extract the rainfall rate profile in a path of some range cells (15 cells). The learning of the network is made using the back-propagation algorithm and several random profiles characterize the learning-set. The input to the network is a vector containing the range cell attenuated measurements of ZH, ZDR while the output is the estimated profile of rainfall rate. In this way a global compensation of the attenuation is realised. Finally will be reported a comparison with the previous compensation technique in the absence and in the presence of receiver noise. In the former case, neural networks allow better performance than parametrizations, in the latter case, the neural networks are very sensitive to the noise when S/N is lower of 25 dB, as well as parametrizations are.
Poster Session 16, Quantitative Rainfall—Multi-Parameter Approaches I
Monday, 23 July 2001, 2:00 PM-3:30 PM
Previous paper Next paper