Improving weather radar estimates of rainfall using feed-forward neural networks

In this paper an approach is described to improve weather radar estimates of rainfall based on a neural network technique. Other than rain gauges which measure the rain rate RR directly on the ground, the weather radar measures the reflectivity ZZ aloft and the rain rate has to be determined over a Z–RZ–R relationship. Besides the fact that the rain rate has to be estimated from the reflectivity many other sources of possible errors are inherent to the radar system. In other words the radar measurements contain an amount of observation noise which makes it a demanding task to train the network properly. A feed-forward neural network with ZZ values as input vector was trained to predict the rain rate RR on the ground. The results indicate that the model is able to generalize and the determined input–output relationship is also representative for other sites nearby with similar conditions.