Determination of Wind from Radar Wind Profilers in the Presence of Bird Clutter Using Machine Learning

Abstract:

Radar Wind Profilers (RWP) play an important role in measuring tropospheric winds. During bird migration seasons, the wind profile data can be severely affected due to bird contamination (clutter). Current wind processing techniques are capable of predicting accurate wind velocities, but they tend to perform poorly in the presence of severe bird clutter. This situation often leads to erroneous or missing wind data. We have studied if machine learning techniques can be used to retrieve wind from bird contaminated measurements. The project was carried out in two phases, using measurements obtained by the RWP operated by the Federal Office of Meteorology and Climatology, Payerne, Switzerland. In the first phase of our study, a Convolutional Neural Network (CNN) was implemented to distinguish bird contaminated data from uncontaminated data. The CNN model performed with an accuracy of 87%. The classified data was then used for phase two of the project where another CNN model was implemented for a regression task. The regression task was divided into several experiments in which the performance of the CNN model was assessed by training the model using different kinds of RWP samples. The results obtained showed that the CNN model was capable of determining radial wind values with a mean absolute error of 0.6 m/s from samples contaminated with bird clutter. These results suggest that machine learning techniques could be used to overcome the barriers faced by the current signal processing techniques in the presence of bird contamination in RWP measurements.