This paper applies statistical modelling techniques (combining a neural network model with seasonal trend analysis) to a comprehensive dataset of soil water contents and pressure potentials measured at 10 cm depth in lysimeters moved to two different locations, in order to identify shifts in hydrological responses to climate forcing.

The shifts in these “in situ” water retention curves (WRC) are intriguing and really quite dramatic (e.g. figures 4, 7 and 8). But I do wonder about the mechanisms and underlying processes. The authors are rather vague about the causes, suggesting that they are due to changes in soil structure tiggered by climate (lines 674-676). I’m not fully convinced about this interpretation, not least because the largest changes seem to occur in the very dry range of the WRC where structure should not play such a large role.

(in this respect, I think the WRC curves should be plotted with matric potential on a log-axis for improved readability. On a linear scale, we can’t really see what is happening close to saturation, which is where most of the structural changes would be expected).

There may be alternative explanations for the observations, including (slowly reversible) swell-shrink behaviour and preferential (non-equilibrium) flow. I would encourage the authors to try to strengthen the discussion and interpretation of the data with respect to the underlying mechanisms, including the above-mentioned processes. Nevertheless, although the responses to climate of apparent WRC observed by the authors “in situ” seem stronger than I would expect (especially in the dry range), I am aware of two previous large-scale (regional-continental) statistical analyses of water retention curves measured in the laboratory that have shown significant  impacts of climatic factors on the structural pore space (Hirmas, D. et al. 2018, Nature 561, 100-103; Klöffel, T., et al., 2024. Geoderma, 442, 116772). These studies could be mentioned as they would give support to the authors’ inferences and interpretations.  

Specific comments

1. Line 40: “indicator” rather than “factor”?
2. Lines 40-41: “Dominant”? This seems a bit far-fetched. I don’t think soil water content at 10 cm depth is the best hydrological indicator of soil health. It could certainly indirectly reflect risks of surface runoff, but then surely surface runoff itself would be a much better indicator of soil health? Contrary to claims made in the paper, I don’t think water content so close to the soil surface says so much about the supply of water to plants (rooting depth and water availability in deeper soil layers will be far more important). I think the authors should tone down the emphasis in the introduction of the relevance of this kind of study to soil health.
3. Lines 42-43: these three papers are not cited in the reference list
4. Line 50: Or (2020) is not in the reference list
5. Line 51: I don’t understand how a modelling approach can be destructive? This should be clarified.
6. Line 77: I think you could also cite Robinson et al. (2019) here (Global Change Biology, 25, 1895-1904).
7. Lines 79-81: It may not be related to only soil structure. What about wettability (hydrophobicity)? Could that also play a role at these time scales? Perennial vegetation can also adapt to climate change. One example comes from an earlier modelling study based on the TERENO SoilCan data (see Jarvis et al. 2022. Hydrology and Earth System Sciences, 26, 2277-2299).
8. Lines 92-93: this statement is disproved by the following sentence (and also by the author’s work). I think this sentence should be deleted (nothing important is lost).
9. Line 93: The authors of this paper are incorrectly specified in the reference list.
10. Line 113: No need to start a new paragraph here.
11. Lines 250-252: this seems very subjective. Why choose these percentiles and these scores? How were they chosen, presumably by trial and error?
12. Line 371: write “water contents” rather than “values” (if I understood correctly)
13. Line 425: “differently” not “different”
14. Line 426: this seems too speculative. You don’t know that it is related to structure. Maybe you could write “….which may indicate”
15. Line 444: “that” not “who”
16. Lines 582-586: This is interesting, but the authors neglect the fact that WRC’s measured in the laboratory do show strong correlations with texture across large regions showing strong climate contrasts. This empirical support for texture-based estimates of the WRC is really very strong, especially in the dry region of the WRC. How can you reconcile your results with this past experience and knowledge?
17. Line 600: this is much too speculative. You need to write this more carefully (“may have”). There are other more plausible interpretations. It could simply be that hydraulic conductivity at and close to saturation has increased. Near-surface water contents during rainfall events of any given intensity would then be smaller. It could also be a consequence of “by-pass” water flow in shrinkage cracks.
18. Line 636: what is meant by “functional integrity”?
19. Lines 643-645: yes, especially if they are statistical models. Isn’t it reasonable to expect that physics-based models might perform better outside the training data?
20. Lines 650: The extent to which the subsoil is exploited by roots should be much more important for plant water supply than the surface 10 cm.
21. Minus signs (for matric potential) are missing on all the WRC plots

In this paper the authors use a neural network trained on meteorological variables, a seasonal trend analysis, and categorical definitions of soil conditions to quantify and evaluate soil water response across soil types and climate regions after an extremely hot summer in Germany.

Overall, I found the study and content of this paper compelling as well as the author’s motivations for incorporating increasingly available remotely sensed products into soil stability analyses. However, I would like to see more discussion of the evaluation of the developed model and reasoning behind decisions, such as using Nash–Sutcliffe efficiency. 

Further, as the authors identify, performance struggles at the site on which the model was not trained. This paper could be made far more robust through the reverse analysis—training the model in wetter conditions—to see how comparable the results are. Or training the model on both a wet- and dry-regime lysometer simultaneously. 

Specific comments

• In Line 75, I’d like to see additional citations to back up the claim of ‘experimental studies’, implying multiple, showing these shifts in soil water content dynamics following extreme events. Or rephrase.  
• Clarifying question, beginning at the end of  Line 79, are you stating that soil texture remains constant through land-use changes? Or in your modeling approach? Or something else?  
• Line 118: While the authors state that predictive modeling providing accurate predictions of soil water dynamics is "not the focus in this study,” I do think that claim requires substantial citations. Alternatively, they could discuss their choice to use the NSE classification here as well as add more information about the pros and cons of the NSE method as this is missing and needed, when the strength of the paper relies on NSE performing as described. Additionally, this sentence (Line 115-119) is quite unwieldy, and it isn’t clear what point the authors are trying to make. 
• Figure 4: Considering the study focuses on the summer of 2018, would highlighting that data be evocative rather than the other groupings? It is not clear why the different time periods have been selected as they are not all the same length. Is this arbitrary or is it related to the mean bias results? Regardless, I think some minor explanation is needed. If the purpose of Figure 4 is only to motivate the selection of a particular lysemeter, the spread of values in the contrasting lysimeter is more than enough, without the additional colors). 
• Line 252: This may be a naive comment, but is there a reasoning behind the numerical values chosen to represent each categorical condition? For example, why is ‘wet’ 1 rather than 10? 
• Relatedly, on what basis are you defining wet, moderate, and dry using the 30th and 70th percentiles? Why not 25th and 75th? Is this arbitrary? Further, have you considered looking at ‘extremely’ wet or dry categories? Considering you call 2018 an ‘extreme’ year throughout the manuscript, I’d be curious to know how incorporating the 95th or 99th percentiles for drying and wetting would be reflected. 
• In the mean bias aggregation, did you also look at a seasonal (wet v dry season) mean bias, or only annual? I would be curious if the response function is stable across seasons as well (to ensure there isn’t variability between wet and dry that are cancelling each other out, for instance). This seems to be most evident in Figure 7, where the predicted water content does not capture both the driest and wettest water content periods. In later years the mean bias improves, but the prediction only appears to have a slightly better fit. 
• Line 423: How do you conclude that this is a result for coarse soil rather than all/many/some compositions of topsoil? When you say ‘this coarse soil’ are you simply referring to that translocated lysimeter? Or making a statement about coarse soils in general? If the latter, I don’t think this is adequately supported.  
• I’d recommend reviewing the language in section 3.2 for clarity, when discussing transposed soil and ‘changed’ climate states. For example in line 447 the use of ‘unchanged climate’ is confusing. Perhaps consider consistent language such as ‘original’ climate, rather than (un)changed to avoid confusion. 
• I would caution against describing the differences in climate between the sites as ‘climatic shifts’ (e.g., Line 518). To me this implies a step change in the climate at a given location, not transposing soil between two sites. The timescale of translocating soil is more abrupt than most/all climatic natural and/or anthropogenic climate change-forced shifts.  
• Were additional training methods explored for this analysis? (e.g., Is one lysometer enough to train this model? Does fit improve or plateau when trained on multiple lysometers? Would training the model on lysometers from each climate perform better than one lysometer from a single climate?)
• Line 600: what evidence for the physical soil response do you have? I don’t see how this could have been determined from the model or information presented prior. Was the autumn and winter of 2019 comparable to conditions in prior years? Would we necessarily expect a rapid rebound to previous year’s conditions? 
• Line 650: I would not classify the top 10cm of soil moisture evaluated in this paper as defining water availability to plants. Most plants have root systems far deeper than 10cm. It seems a stretch to connect these results to agricultural impacts as surface vegetation was not varied or evaluated. 
• 673: ‘extreme climatic events’ seems overly broad, as you’re only focusing on a single intense drought, you cannot draw conclusions against other extreme events (e.g. extreme precipitation).
• 675: how are you defining ‘past’ here? Did you evaluate how unique the 2018 drought was in these locations/for these soils? 

Technical corrections

• Inconsistent spacing in em-dash usage. There should not be any spaces between words separated by em-dashes (e.g. use is correct on line 22, but not correct on lines 54, 55, 77, 78. 
• Line 75: possibly missing an ‘a’ or time scale should be plural. 
• Line 101-102: ‘for a set of lysimeters’ I’m not sure ‘for’ is the correct preposition, perhaps ‘with a set of lysimeters’? Unless I’m misunderstanding the function. 
• Line 114: for clarity, I’d recommend adding commas around ‘for example’ to set off the clause. 
• Line 116: appears to be an extra space between ‘which’ and ‘the’ 
• Line 160: I would recommend offsetting the place names with commas (e.g., “including both, a wetter site, Selhausen, and a drier site, Bad Lauchstädt, …”). Alternatively, you could use parenthesis. 
• Please align tense, particularly in the methods section, where both present and past tense alternate quite frequently. 
• Line 215 is stating the same information as Line 130. I recognize the authors are re-introducing the reader to the input variables, but I would recommend restructuring as to not imply that water balance has not been discussed previously. My suggestion would be to either re-introduce all inputs in this section or, better yet, move this discussion up near Line 130. Moving the input variable discussion higher up would also reduce a bit of confusion with the title of subsection 2.2.1 as you aren’t discussing the STD until the second paragraph. In that same vein, the beginning of the 2.2.1 subsection implies that the input variables were selected after data cleaning, which does not seem to be the case (as the input variables were selected for cleaning/alignment). 
• Line 298: Should the Adam optimizer be cited? 
• Line 322, 670: I’m not sure the 2 hyphen is needed in Nash–Sutcliffe efficiency, nor should the E be capitalized. It may also be more clear to state: ‘we use the Nash-Sutcliffe efficiency coefficient (NSE, see eq. 5) as a general descriptor…”
• Line 444: use ‘which’ or ‘that’ rather than ‘who’
• Lines 519, 520: hyphens used instead of em-dashes. 
• Line 615: missing a period. 
• Lines 639, 641: hyphens used instead of em-dashes.
• Line 669, 672: the preceding bullets end with a period, whereas these two do not. 
• Line 676, 676: include only one does before the (i) and make require plural.