Please see attached.

The paper demonstrates the use of random forest regression (RSR) to estimate wind speed profiles based on satellite 10m winds and ERA5 temperature data (SST and 2m). The work is novel and the comparison is quite extensive comparing with several independent sources. My main concern is that the authors compare their results to a neutral log profile and a power law profile with a fixed exponent. The neutral log law is used as the authors state that there is insufficient information to make the stability correction using a diabatic profile. This seems somewhat surprising. They are using ERA5 as an input to their RSR model which already uses delta T. This could equally be used along with wind speed to infer a bulk Richardson number which could be related to Obukhov length (L) and a formulation of the psi(z/L) function such as that by Businger and Dyer used to make the stability correction to the neutral log law. Indeed other parameters from ERA5 (e.g. sensible heat flux) could be used to get a more accurate estimate of surface stability. This would seem to be a much fairer comparison than using the neutral log law. Furthermore, it would be interesting to see a more detailed assessment of the accuracy of the RSR/log law/power law profile as a function of direction to see if the coastal transition plays a role in model accuracy. Finally, the results of the performance metrics for the floating lidar sites would be more readable as a table as was done for the ASOW sites.

Minor comments:

• The value of z0 for the log law extrapolation does not seem to have been mentioned (unless I miss it somewhere). Or was a Charnock relationship used?
• Units are missing in Table 4.
• On page 8, the word 'importances' is used several times. This sounds odd and I suggest that 'importance' is used as a collective noun.
• Line 522: change 'shown decrease' to 'show a decrease'

Please find comments in the attachment.