|The authors have carefully revised their manuscript which has substantially improved and is almost ready for acceptance in my opinion. The additional work performed strengthens the conclusions of the article. However, some issues remain that should be addressed by the authors. Some of them have already been mentioned in my first review. I thus just copy in my former comment that I would like the authors to answer:|
“A major problem of the paper as it stands now is, in my opinion, the lack of validation with independent measurements: The training data set for 32 glaciers is based on “remote-sensing” (Rabatel et al., 2016). As this lies the foundation to the entire study, more effort should be invested to describe this data set (methods, uncertainties). The training data set also contains models and assumptions – the annual SMB of these glaciers has not been directly measured. This needs to be emphasized. Measured (!) information on SMB is available from two sources: (1) the direct glaciological surveys, (2) geodetic surveys.
Although (1) is probably included in the training data set, it should explicitly be shown (e.g. figure with cumulative SMBs) how well the DL approach reproduces the observed SMB series. This would give direct evidence on the performance of the approach, independent from the training data set by Rabatel et al. (2016) that also includes model assumptions.”
The description of the basis of the principal training data set (Rabatel et al., 2016) is still very limited (see page 3, line 16). The reader just knows that it is based on “remote sensing”. However, it is clear that remote sensing alone is (yet) unable to provide annual glacier-wide mass balance. It should be described which approaches and assumptions went into the training data set and that glacier-wide mass balance is not measured, as in the case of the four GLACIOCLIM series, but is based on modelling optimally constrained with geodetic mass balances and snowline observations. I do not at all want to imply that the data set is unsuitable or inaccurate but it should be honestly stated that these are not actually “observations” as the annual mass balances are referred to several times. Related to this, I would also ask the authors to follow my advice of comparing the annual (!) series from their approach against the four GLACIOCLIM data sets. This is directly measured data and would allow investigating whether DL is able to capture these observed year-to-year variations.
Some additional comments:
- What is shown on top of page 4 is not actually Data. It rather seems to be a description of the method to incorporate them. Please try to better separate data used in the approach from the methods. I realise though that this is not trivial to achieve here.
- A general comment (sorry, for not bringing this up already in the first review): Throughout the paper mass balance is referred to as SMB (surface mass balance). But does this approach actually compute SURFACE mass balance? And not rather the total mass balance of a glacier (see Cogley et al., 2011, for an overview)? The approach is trained on the Rabatel-dataset which depends on geodetic balances for constraining the mass loss. Geodetic mass balance covers both surface mass balance and basal/internal balance. Consequently, I would consistently omit the “surface” when you refer to mass balance. Anyway, this is just terminology, as it is clear that the difference between SMB and total mass balance will be minimal for Alpine glaciers.
- It is a pity that the new validation with independent geodetic mass balances has not made it into the main paper. I think it would strongly support the credibility of the data set and is important enough to not only be presented in the Supplementary.