Extended Abstract (148K)P2B.11
Advances in radial basis function neural network algorithm for radar rainfall estimation
Gang Xu, Colorado State University, Fort Collins, CO; and V. Chandrasekar and W. Li
Adaptive Radial Basis Functions (RBF) neural networks using vertical reflectivity profiles as inputs can be successfully used to estimate ground radar rainfall (Liu et al 2001, Xu et al 2001, Li et al 2002). The training process for adaptive RBF neural networks involves two steps: building the initial RBF model and adaptively updating RBF models (Liu et al 2001). It is a nontrivial task to determine the parameters of the initial neural network such as the number of centers in RBF neural networks and the widths of RBF centers (Liu et al 2001). Along this direction Orr (1998) proposed a method to optimize the width of RBF centers by explicit search. Two principles are important during training neural networks; the first one is to avoid overfitting to specific data and the second one is to tune the neural network for optimal performance. It will be desirable to develop an automatic scheme to build the initial network and progressively update the models to eliminate the manual tuning effort. This paper describes a new algorithm to automatically build the initial RBF model as well as progressively update the models for radar rainfall estimation. The RBF architecture and all internal parameters, including number of centers and widths of RBF centers, can be autonomously determined during the training phase. This algorithm also includes an embedded mechanism to avoid the overfitting during the training. We have applied this algorithm to the ground rainfall estimation from the WSR-88D radar data in 1998. We progressively train and apply RBF models to estimating the daily and hourly rainfall accumulations from vertical profiles of reflectivity and compare them with ground gauge data. Application to one year of WSR-88D radar data over Melbourne, FL shows that this improved radar rainfall neural network can estimate pointwise daily accumulation to a normalized standard error of 25% with negligible bias. Easy tuning procedures are also demonstrated in this paper.
Poster Session 2B, QPE/Climate Poster
Thursday, 7 August 2003, 1:30 PM-3:30 PM
Previous paper Next paper