Dear authors
The reviewers gave a number of good suggestions to improve the quality of the paper and also to clarify some part that in fact could lead to misinterpretations. It would be important to discuss the importance of uncertainty quantification as suggested by Reviewer 2 and also a possible list of uncertainty sources to consider and their expected role. The points from Reviewer 1 are also very important in order to make the paper clear. The possibility that re-implementation of codes could lead to (even large) errors must be clearly discussed. I agree with the author that the check and analysis of the differences among implementations would require an effort that if out of the scope of the paper but for the same reason it is important to clearly highlight the risks and possible effect on the results. This should include also the risk that the original data version used in the different analysis (OzFlux, DINGO, FLUXNET2015) could be not exactly the same. In order to help this analysis and discussion I agree with the suggestion from the authors of an extra-figure focusing on a sub-sample of sites where more in-depth analysis are done, for example comparing the ustar thresholds calculated in the different processing methods in order to check if the differences are there (the thresholds used and their uncertainty are available in the FLUXNET 2015 dataset). However I don’t agree with the sites proposed by the author (AU-Cpr, AU-DaS and AU-Whr): the one where larger differences are found should be checked, first of all AU-Tum (even if already somehow analysed in the text).
For the DINGO methodology, please cite the BGD paper if you don’t’ want to integrate the description in your manuscript.
Two additional notes from me:
- Double counting is an issue only when wrongly applied, i.e. when applied before the storage correction (where storage is relevant) (page 13 line 30).
- The FLUXNET processing uses a combination of two method for ustar calculation… for this reason it is not clear how you could conclude that the same CPD method in FLUXNET calculated 0.24 as threshold… in fact the threshold calculated by the CPD in FLUXNET (using the original Barr et al code) for 2013 was 0.51 but the uncertainty in the threshold was large: 25 percentile was 0.19, 75 percentile 0.55. Even worst if you look at the 5-95 percentiles: 0.09 – 0.68. This is the reason why I agree with Reviewer 1 that such comparisons are risky and must be really well described, with caution on the conclusions…

On all the other points it looks that you and the reviewers already agreed on the modifications.
Ciao
Dario

Dear authors

thanks for the new version that in my opinion is improved and correctly addressed all the points raised from the reviewers. I only suggest to you two minor points that you are free to decide to add or not, for the rest the paper is ready to get published for me.

1) it seems that in your processing you filter for ustar only night-time data below the threshold (page 17 line 19). If so consider that in the FLUXNET 2015 also daytime data below the threshold are removed and also the first data point after a period of low turbulence. This could also contribute to explain the differences.

2) in the Uncertainty paragraph it could be also important to refer to the footprint and the possibility that in some sites it extends outside the target ecosystem.