The manuscript introduces a novel, largely automated structure-from-motion (SfM) photogrammetric system for high-resolution monitoring of soil surface change over 3.5 years on agricultural hillslopes. Synchronized DSLR cameras (triggered by rainfall events and timers) capture daily imagery, which a custom Python workflow processes: it time-synchronizes photos, applies a convolutional neural network to detect ground control points under varying conditions, and runs Agisoft Metashape SfM to reconstruct daily 3D soil-surface point clouds. From these, daily digital surface models and change-of-surface (DoD) maps are derived at millimeter-scale resolution. The method is validated against terrestrial laser scanning (TLS) and UAV photogrammetry. The data from a freshly tilled loess field demonstrates detailed topographic changes following tillage and rainfall. Overall, this approach is innovative and promising for tracking erosion dynamics at high spatial and temporal resolution.

Major Comments

• The manuscript is very comprehensive, but some methodological sections (e.g. the multi-step GCP tracking and neural network architecture) are very detailed. To improve readability, consider moving deep technical specifics (such as the exact CNN architecture, clustering parameters, or algorithmic steps) to an appendix or supplementary material. This would shorten the main text while still making the detailed methods available for interested readers.
• The authors claim that “the resulting high-resolution datasets are valuable for analysing erosion dynamics and validating soil erosion models.” However, the study is primarily a methodological demonstration rather than an explicit evaluation of model validation. The conclusions should be tempered to reflect this. If model validation is to be a major claim, the manuscript needs to either provide analyses that link the data directly to model performance or clearly state that testing such models is outside the current scope.
• A critical limitation is that the system records all surface elevation changes, including non-erosional effects (e.g. soil compaction or tillage settling). The observed strong negative correlation between rainfall and elevation loss in the first week post-tillage indicates both erosion and compaction. However, without a method to distinguish these processes, the results cannot be interpreted purely as erosion. The authors note this challenge but provide limited quantitative separation techniques. The referenced approach by Epple et al. (2025) has not been applied at this temporal scale. The authors should discuss and, if possible, implement methods to discriminate erosional changes (e.g. mass removal) from other changes. For example, integration with sediment traps or other measurements (see below) could help verify when soil material has actually been removed by erosion versus simply redistributed or compacted.
• The monitoring system achieved data on only 55–69% of survey days. Hardware issues (e.g. rigs knocked over by storms) and maintenance downtime caused significant gaps. This undermines the claim of “near-continuous” monitoring. The manuscript should acknowledge this limitation more explicitly. In particular, discuss how the 31–45% data loss affects the claimed continuity, and consider suggesting improvements for hardware robustness or system redundancy. The data gaps also impact the applicability for validating models like RUSLE that rely on annual or longer-term averages; this should be discussed. For example, if many events are missed, event-based model validation is compromised.
• The system’s accuracy varies spatially, with errors growing toward the edges of the region of interest (5–15 mm). Such spatial bias can systematically affect erosion calculations, especially for analyses covering the whole area. The authors should discuss this limitation, for instance by quantifying how much area is affected by larger errors and whether analysis should be restricted to a smaller central zone for reliable measurements. Additionally, providing confidence intervals or error bars (rather than just mean errors) would clarify how spatial uncertainty influences change detection.
• GCPs were noted to move due to animals, tillage, and storms. It is unclear how these movements propagate into final surface uncertainty. The manuscript should quantify or at least discuss the impact of GCP displacement on accuracy. For example, how often did GCPs shift beyond measurement error, and how does that translate into potential vertical error in the DSMs?
• High-resolution temporal data are more suitable for some erosion models than others. The discussion should explicitly address which types of models can benefit. For instance, process-based models that use detailed spatial inputs (such as RillGrow or similar rill erosion models) can leverage sub-daily or event-scale data. In contrast, simple empirical models (e.g. RUSLE) operate on annual average erosion rates and cannot directly utilize such fine temporal detail. The authors should clarify that the proposed monitoring approach is especially useful for validating physically based, high-resolution models, and acknowledge the limited benefit for annual-scale models. Explicit examples (like RillGrow vs. RUSLE) would illustrate this point.
• The innovative multi-camera synchronization needs clearer explanation. The temporal drift correction procedure, in particular, should be described in more detail to ensure reproducibility. It would also help to state whether camera internal calibration was performed only once before the campaign or periodically (e.g. after storms), and whether any re-calibration was necessary during the study. Details on camera maintenance and calibration will improve confidence in the results.
• The use of a 3-sigma filter to remove outliers may be too aggressive for detecting subtle erosion features; the authors should justify this choice or consider less aggressive filtering. Additionally, the chosen 5 mm grid resolution for DSM interpolation should be justified: is this based on camera resolution, expected soil roughness scale, or processing considerations? Discussing why 5 mm is appropriate (and whether coarser or finer resolutions were tested) would strengthen the methods.
• Accuracy is currently given in millimeters, which is precise but not easily interpretable in terms of soil loss or erosion impact. The authors should consider converting key error metrics into more meaningful units (e.g. tonnes per hectare per year) or comparing them to typical erosion rates. Furthermore, statistical details are needed: provide confidence intervals on accuracy estimates, and separate systematic versus random errors if possible. An analysis of how accuracy depends on surface roughness or vegetation cover would also be valuable (since rougher surfaces may degrade accuracy).
• Where possible, compare the SfM-based erosion estimates to traditional measurement methods (e.g. sediment traps, erosion pins) to demonstrate practical utility. This could also help in distinguishing erosional vs. non-erosional volume changes.
• The study area was a bare, conventionally tilled plot, which simplified reconstruction. Real agricultural fields often have growing vegetation, which can complicate SfM. The authors mention using a machine-learning method to filter vegetation in some cases. It would strengthen the paper to discuss how the system could be adapted for vegetated conditions (for example, using spectral filters, morphological editing, or annual timing). The transferability claims should be more specific: what additional challenges or modifications would other sites (with different climate, terrain, or vegetation) require?
• A valuable addition would be to suggest or demonstrate integration with other erosion measurements. For example, combining surface change data with sediment trap records or runoff measurements could help distinguish erosion from compaction. If changes in the soil surface do not correspond to transported sediment, one could infer non-erosional processes. The manuscript should at least discuss this possibility and how future work could integrate multiple datasets.

Minor Comments

Line numbering should be continuous throughout for ease of review reference.

Some abbreviations are introduced without definition (e.g. RTC, IoT, LoD, M3C2). Define all acronyms at first use.

The discussion of transferability would benefit from concrete guidance: for instance, recommended mounting improvements (e.g. sturdier rigs, solar power redundancy), or software alternatives (since Agisoft Metashape is proprietary, the authors might suggest open-source SfM tools for reproducibility).

In this paper, a new approach for an almost-continuous monitoring of erosion at the hillslope scale is presented. The idea of conducting a continuous survey is very innovative and challenging, and the methods presented are very promising. However, the paper in its present form presents some criticisms and necessitates some revisions.

• In the “abstract”, a lot of acronyms (SfM, GCP, CNN, DoD, LoD) are presented. However, most of them are not useful for the “abstract” section and could be removed.
• The “introduction” section is too simplistic. At least, the papers dealing with the use of SfM for monitoring hillslope erosion at the event/run scale should be presented.
• Line 72: The start of the sentence is a bit twisted. Please, revise.
• Line 77: It is not clear if the investigated field was hydraulically delimited or not. This implies significant differences in terms of runoff generation and sediment transport dynamics.
• Line 79: I think that the “e” should be deleted.
• Some acronyms are not defined in the text.
• The sequence of tenses is not always optimal. Please, revise the text.
• Lines 101-102: The sentence is quite confusing, and some words are repeated.
• The quality of the legends in Figure 3 is too low.
• Line 432: there is a point after mm that should be deleted.
• How do you discern the elevation changes due to vegetation or post-tillage settlement from those due to erosion and deposition phenomena? I believe that resolving this aspect is crucial for the satisfactory application of the presented methodology. Even if briefly discussed, it remains incomplete.
• Have you thought about ways to increase the percentage of total usable days? Is it possible to further protect the setup and avoid big gaps in data collection?