In light of the difficulty in finding referees, I have taken over the review process. As a note about this: I have significant background and expertise in geomorphology and Earth-surface processes, some in subsurface structure and sedimentary geology, and none in machine learning. Because neither of the two first-round reviewers raised any major concerns regarding the technical details of your ML approach, I am proceeding under the assumption that this is fine, and will not be commenting on it here.

First, I appreciate your responses to the reviewers and improvements to the manuscript. These are substantial and a major step forwards.

Second, based on my reading of Reviewer 2's comments and your responses, I believe that you addressed some of the issues, but that there remains the matter of a more fundamental disconnect between your approach here and questions regarding the broad applicability of your method and the geological reality, especially in more complex systems. I will enumerate some specific issues here in hopes that you can clarify, qualify, or modify your work in such a way to address these concerns of mine, some of which were inspired by Reviewer 2 and your response, and some of which are new but not entirely unrelated. If you can address these two concerns, then I think that we will be able to help you move your paper forward towards publication.

\begin{itemize}
\item \textbf{Generalizability.} Your study site has been quite tectonically stable for a very long time, and was never glaciated. Therefore, the geomorphic evolution is straightforward: erosoin of bedrock uplands produces sediments that accumulate in basins, and geological and climatic processes lead to the formation and filling of valleys. In this case, these valleys have been loci of deposition throughout the Cenozoic. However, Reviewer 2's comment notes that your situation does not match that of many other parts of the world, where surface topography and subsurface structure are not linked. As an example that I know well, note Minnesota (USA)'s depth to bedrock (\url{https://mngs-umn.opendata.arcgis.com/} -- scroll down), topography (http://arcgis.dnr.state.mn.us/maps/mntopo/), and aeromagnetic anomalies (https://mngs-umn.opendata.arcgis.com/app/the-aeromagnetic-database). You would not obtain the same answer with your methods here, where I am! However, as you note, it works for both your training and test regions, both of which are within the same physiographic and geological region. Therefore, I think that you need to couch your algorithm as an approach to find and characterize valley fill when topography is still present. And in so doing, it might be good to look into the work of Mey et al. (2015: \url{https://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/2014JF003270}, who also address this problem.
\item \textbf{Fitting geological reality or a geophysical inversion?} I am concerned about your note regarding the valley structure being complex and discontinuous. This is inconsistent with your description of fluvial deposits, which should form a continuous sedimentary package (i.e., streams flow consistently downhill). There are two options. One is tectonics and localized deformation that reshaped the contact between the basement rock and the overlying fluvial deposits. The other option, which I find to be more compelling, is that the geophysical inversions may be imperfect. Although I am not as familiar with aeromagnetic data, I believe that inverse-distance effects would cause narrow sections of the paleovalley to be harder to detect and/or to seem to terminate at a shallower depth than wider portions of the valley. If I am correct, then your ML-based fit would be tuned to the details in a depth-and-wavelength-sensitive geophyscial inversion rather than designed to match geological reality based on lithological data and interpretations.
\end{itemize}


Line-by-line comments:

\begin{enumerate}
\item[12.] tomography $\rightarrow$ topography?
\item[75.] their $\rightarrow$ its
\item[79.] "sandplain" is not a genetic term. Later in the paper, it seems that you indicate this to be aeolian. You should make this clear here, because it importantly indicates that you are inverting across a buried and preserved paleolandscape and its deposits. (You can expand upon this to in regards to the geological setting.)
\item[82.] This is the core of my second major point, above.
\item[85-87.] See the first point; I think that this might not be as generally applicable as you state, especially since you test and validate it in the same stable geological setting.
\item[90-91.] This indeed is THE key limitation, and I think it deserves to be more clearly stated, as this will help readers at once recognize whether your approach is one that they could use in their system or not.
\item[103, Fig. 1, \&c.] Is your use of "image" standard? Becuase to me, these are ``arrays" or ``layers of inversions", but not actually images.
\item[154.] This is your first mention of a CSIRO data set. What is it?
\item[157.] I assume that this is the ``sandplain'', per my above comment.
\item[159.] The MODERN valley bottom. In addition, you have not yet introduced the valley-bottom flatness index. It seems that this paragraph may be out of place, and should follow the paragraph below in order to provide the requisite information first.
\item[169.] Define how the MrVBF is calculated and the details of the source DEM (which data source, resolution, etc.)
\item[176.] MrVBF is a derived data product, not a data source in itself.
\item[183.] Not every deposit filling a valley is an aquifer. Is this formula sensitive to the nature of the sedimentary fill? Associated with that, the cross section in Fig. 2B indicates an aquitard between two aquifers.
\item[228.] What is a fully connected layer in the encoder? Could you provide a bit more ML background to help the reader to understand why it is important and creating the observed result?
\item[287-289.] I appreciate your desire to contribute, but do not think that your tool has generic applicability based on the points that I have raised.
\end{enumerate}

My thanks for your patience in this year.

I have carefully reviewed your responses and all changes. I have just some very minor comments, which do not require any further review by me. Please submit these small/technical corrections based on my comments below. I am happy to accept your paper for publication, and congratulate you at the end of a long road.


l 91-92: "It is also worthy of note that the topography of the study area is very subdued, with contemporary draining systems being discordant with respect to their ancient precursors.": I had thought that the key to your work was the fact that the topoography, while subdued, was concordant with the paleovalley systems.

l 98: "Ma" alone -- "ago" not needed.

l 108: The existence of plays may bolster your counterargument to my last review: either sediment deposition or neotectonics must have split the valley into internally drained basins by this time.

l 135-139: You state that the geophysical data match geological reality, but what would really be good is to cite sources (e.g., with borehole logs) to indicate this.

Fig. 1: This is an excellent addition.

l 252: I would suggest formatting this as a full reference to go in the reference list, especially considering that it is a doi. If it stays a link, the Copernicus staff will ask you to give the date of last access.

l 384-385: Not sure if this sentence is needed; the prior sentence states the same, but more precisely.