A NEURAL-NETWORK APPROACH TO ESTIMATING RAINFALL FROM SPACEBORNE MICROWAVE DATA

Rainfall is a key parameter in the study of global climate budget and climate change, Various techniques use microwave (MW) brightness tempe rature (BT) data, obtained from remote sensing orbiting platforms, to calculate rain rates, The most commonly used techniques are based on r egressions or other statistical methods, An emerging tool in rainfall estimation using satellite data is artificial neural networks (NN's), NN's are mathematical models that are capable of learning complex rela tionships, They consist of highly interconnected, interactive data pro cessing units, NN's are implemented in this study to estimate rainfall , and backpropagation is used as a learning scheme, The inputs for the training phase are BT's and the outputs are rainfall rates, all gener ated by three dimensional (3-D) simulations based on a 3-D stochastic, space-time rainfall model, and a 3-D radiative transfer model, Once t raining is complete the NN's are presented with multi-frequency and po larized (horizontal and vertical) BT data, obtained from the Special S ensor Microwave/Imager (SSM/I) instrument onboard the F10 and F11 pola r-orbiting meteorological satellites, Bence, rainrates corresponding t o real BT measurements are generated, The rainfall rates are also esti mated using a log-linear regression model, Comparison of the two appro aches, using simulated data, shows that the NN can represent more accu rately the underlying relationship between BT and rainrate than the re gression model, Comparison of the rates, estimated by both methods, wi th radar-estimated rainrates shows that NN's outperform the regression model, This study demonstrates the great potential of NN's in estimat ing rainfall from remotely sensed data.