Dear Lee,

I read your paper with great interest and agree with both reviewers that this work has great potential to evaluate to which extend machine learning methods can help to classify soil moisture events and to search for underlying reasons. Yet, I think in line with the assessment of reviewer 2 that the manuscript needs major revisions to fully explore it’s potential. In addition to the points raised by the reviewer, I I’d like to bring the following points to your attention.

Figure 2 is indeed difficult to digest. Maybe it is better to present boxplots not using the profile names but their topographical elevation on the x-axis (more information), and it might be better to separate different depths into different panels.

I have furthermore problems with the definition of the maximum soil moisture variation VAR. A normalization with the antecedent soil moisture is hydrologically misleading. Let me explain this based on two cases:
1) Theta_max = 0.4, Theta_asm =0.3 gives a var of 33%,
2) Theta_max = 0.15, Theta_as = 0.05 gives a var of 200%,

In fact in the first case the storage change was 200% larger than in the second. I think a normalization with total precipitation depth is hydrologically much more appropriate. You could calculate storage changes, for instance by multiplying changes in theta with the respective depth increments that are represented by these measurements. VAR will then compare normalized storage changes, and the normalization is constant and not variable due to the variability of ASM between the profiles (see Fig. 3). To me this is a hydrologically much more meaningful index, which relates storage changes to the total forcing.

Related to this, I think that Fig. 3b is misleading, as the variations downslope are normalized with systematically elevated antecedent soil moistures. The proposed normalization with total precipitation will rectify this and the downslope var will become likely larger than the upslope var. And it might change the results of the clustering and will differences more meaningful.

Figure 4 is interesting, but the numbers are too small. Please clarify whether this is the coefficient of determination R2 or the correlation R. The corresponding text mentions both. I also wonder whether you could correlate ASM and optional other predictors with VAR (to explain variations between the profiles).

Last but not least, I agree with reviewer 2 that a clear physical interpretation of the clusters is highly desirable. Boxplots are an abstract representation of an event. In a different context (model error classification), Reusser et al. (2009) used typical errors in runoff events and classified those with am SOM. I am sorry to mention my own work in this context, this approach helped to interpret the corresponding error clusters (and to create an imagination how they looks like). Another option would be to show the events closed to the cluster centroids.

Best regards,

Erwin Zehe

References:
Reusser, D. E., Blume, T., Schaefli, B., and Zehe, E.: Analysing the temporal dynamics of model performance for hydrological models, Hydrology And Earth System Sciences, 13, 999-1018, 2009.

Dear Dr. Lee,


I am very much pleased by the way how you addressed the reviewers and my recommendations in the revised manuscript. So I don’t see any reasons to further hold this work back from publication.

Best regards,

Erwin Zehe