A neural network based method to adjust weather radar estimates of rainfall to rain gauge measurements using the vertical reflectivity profile

Extensive analysis of rain gauge and weather radar data in Austria showed that the metering precision of the radar largely depends upon the type of rainfall. On the one hand the radar underestimates low-level advective rainfall because of shielding effects and inhomogeneous beam filling especially over complex terrain. On the other hand very strong echoes of shower events are often caused by hail and lead to an overestimation of rainfall depths by the weather radar. Therefore the adjustment methodology has to take into account the type of rainfall present. The vertical profile of reflectivity (VRP) can provide this information. It is gained from volumetric data acquired by the C-band weather radar station on Mt. Zirbitzkogel, Austria. This paper presents a method to use the VRP over rain gauge sites to train a neural network to predict the rainfall depth measured at ground level. Based on long-term weather radar and rain gauge archive data, a great variety of rainfall events different in type and intensity formed the model and led to stable conditions. First results showed that the neural network model can correct for a large part of the errors inherent to the radar system and that the model is also representative for sites with similar elevation and distance from the radar. The thus adjusted radar data provide the basic input of an already existing rainfall-runoff model using original radar data so far.