Radial wind velocity retrieval from Doppler radar and lidar measurements using Deep Learning
A Pulse-Doppler Radar Wind Profiler (RWP) is an active remote sensing instrument used in meteorology whose output product is a 3D wind field. A new method to retrieve radial wind velocity from spectrograms using Convolutional Neural Networks is introduced. The collection of data is provided by the Federal Office of Meteorology and Climatology MeteoSwiss and collected in Payerne, Switzerland. It covers the summer months of the year 2020. As a first step, spectrograms are split into different classes as follows: (0) no visible/ measured wind, (1) visible wind only, (2) massive bird contamination (3) slight contamination and wind still visible. In terms of accuracy, precision and recall, the model achieves a solid performance of 94% on the test set with a tendency to mix classes 1 and 2 with class 3. Spectrograms either too heavily contaminated (class 2) or lacking a wind signal (class 0) are discarded in a second phase. A Doppler lidar provides the radial velocity shift for each spectrogram. Accross the test set, a R2 of 0.97 is obtained along with a mean absolute error of 0.22 ms−1. The study undoubtedly opens up a new range of possibilities regarding the processing of wind profiler measurements. Modern techniques have been successfully deployed, and coupling them with the existing and approved algorithms might strengthen the output product robustness.
ROMERO GRASS_PDM PRINTEMPS 2020.pdf
Publisher's version
restricted
Copyright
11.45 MB
Adobe PDF
61a6dd0d31beb02cbc513b8f80452d15