The precision of radial velocity ( RV ) measurements depends on the precision attained on the wavelength calibration .
One of the available options is using atmospheric lines as a natural , freely available wavelength reference .
Figueira et al .
( 2010 ) measured the RV of O _ { 2 } lines using HARPS and showed that the scatter was only of \sim 10 m/s over a timescale of 6 yr .
Using a simple but physically motivated empirical model , they demonstrated a precision of 2 m/s , roughly twice the average photon noise contribution .
In this paper we take advantage of a unique opportunity to confirm the sensitivity of the telluric absorption lines RV to different atmospheric and observing conditions : by means of contemporaneous in-situ wind measurements .
This opportunity is a result of the work done during site testing and characterization for the European Extremely Large Telescope ( E-ELT ) .
The HARPS spectrograph was used to monitor telluric standards while contemporaneous atmospheric data was collected using radiosondes .
We quantitatively compare the information recovered by the two independent approaches .

The RV model fitting yielded similar results to that of Figueira et al .
( 2010 ) , with lower wind magnitude values and varied wind direction .
The probes confirmed the average low wind magnitude and suggested that the average wind direction is a function of time as well .
However , these results are affected by large uncertainty bars that probably result from a complex wind structure as a function of height .
The two approaches deliver the same results in what concerns wind magnitude and agree on wind direction when fitting is done in segments of a couple of hours .
Statistical tests show that the model provides a good description of the data on all timescales , being always preferable to not fitting any atmospheric variation .
The smaller the timescale on which the fitting can be performed ( down to a couple of hours ) , the better the description of the real physical parameters .
We conclude then that the two methods deliver compatible results , down to better than 5 m/s and less than twice the estimated photon noise contribution on O _ { 2 } lines RV measurement .
However , we can not rule out that parameters \alpha and \gamma ( dependence on airmass and zero-point , respectively ) have a dependence on time or exhibit some cross-talk with other parameters , an issue suggested by some of the results .