Kuo-lin Hsu, Yang Hong, and Soroosh Sorooshian Center of Hydrometeorology and Remote Sensing Department of Civil and Environmental Engineering University of California, Irvine E-4130 Engineering Gateway, Irvine CA 92697-2175 Tel: 949-824-8826; Fax: 949-824-8831; E-mail: firstname.lastname@example.org
Artificial Neural Networks (ANN) has been found to be very useful in data classification and regression fitting of unknown functions in a noisy environment. Many applications have demonstrated that ANN is capable of predicting environmental variables for which their underlying processes are nonlinear and not clearly known. In this study, a neural network model specially designed for rainfall estimation at fine scales from satellite information, entitled Self-Organizing Nonlinear Output (SONO) map, is presented. This model has been designed for efficient and effective estimation of model parameters and system output. SONO performs two basic functions as “classifier” and “approximator”. As a classifier, the SOFM (Self-Organizing Feature Map) groups selected image features and turns “on” or “off” of nodes in the nonlinear output layer. As an “approximator”, SONO calibrates a nonlinear satellite infrared (IR) and rainfall rate (RR) function for each classified SOFM classification group. SONO combines multiple data sets from radar, microwave, and infrared radiometric sensors to exploit satellite rainfall detection and estimation at fine scales. Comparisons of SONO with other methods, including fixed thresholds and nonlinear exponential function approaches, for the rainfall estimates of the summer monsoon rainfall over the southwestern United States was examined in this study. The results show that the SONO estimates are performing better that that of other methods. Details of the model development and comparison will be presented.