1. GENRAL COMMENTS

• The introduction fails to clearly highlight the scientific gaps that need to be filled.
• There seems to be some information missing from the introduction that would help  to really understand your objectives
• A grain size description of the reference data (manual or pseudo ground truthdata or Wolman data) is missing, it should be provided either in the methodology section or in the results section.
• information about distribution of samples surface area/d_max area or grain number per samples could be provided instead of active chanel width in figure 1B
• a few small details may be missing in the methodology and results section. (see below the listed details by lines)
• the errors concerning the estimation of the high percentiles by the photosieving tool (here DGS) and the significant scatter for the prediction of the high percentile with the roughness calibration could be related to a problem with the surface area sampled by photosieving not being large enough (1m² for certain samples being too small), coupled with a surface area for the extraction of the roughness values which is different from that for the extraction of the pseudo ground-truth D₅₀. Each member of the data pair does not cover the same extent
• options to be implemented to give more weight to this study: 
  • a second photosieving tool of the same type but more recent such as Sedinet (Buscombe 2020) or an oriented object detection tool could have been tested in parallel to DGS to evaluate the importance of the tool used on the subsequent performance of the D₅₀ estimation with roughness on this large dataset composed of 12 rivers. 
  • In the same vein, a second mesh size to extract roughness could have been tested to assess the weight of this parameter.
• The error of the D₅₀ estimate through the roughness method seems to be accurate compared to the Wolman field data. But it is not certain that these differences in D₅₀ in Table 6 characterise well the performance of your method given that such a large spatial integration may include compensation. 

• To evaluate the performance of large-scale spatial estimation of D₅₀ using an RTK-equipped drone (but GCPs are ultimately used, and the errors are not compared with point clouds constructed without the use of RTK during photo acquisition).
• Evaluation for a sedimentological application? (not clear).
• It is the first examples of application of such a method in this field? (in relation to the results obtained in section 3.3.1 and 3.3.3 ). 

• The first one mentioned doesn't seem to be an objective in itself. You are evaluating the specific error of the DSG tool for photosieving. Studies already exist on this subject. In reality, I think you want to correct the ‘systematic’ errors of your automatic photosiving to then ensure that you minimise the errors of your D₅₀~roughness calibrations. You do this because you know from your literature reviews that automatic photosiving methods are biased. But the problem is that you never talk specifically about DGS in your introduction, so it's complicated for you to introduce, and for the reader to understand, your 1st objective. Eventually you could have had this objective if you had compared several photosieving tools (at least two). In the end, I would say that either this is a sub-objective of your objective 2, or it could be an objective 1 called ‘to evaluate the automatics phototosieving error and developpement of correction with field-base calibration’.
• Perhaps objective 2 should be reformulated. It seems to be strongly linked to the use of RTK, but then neither the results nor the discussions refer to RTK.

General comments:

This is a very good manuscript, whose topic is highly appropriate for the readership of ESurf. The introduction is thorough and an excellent review of the state of science in this field/technique to mapping sediment size. The research questions are excellent (L105-111) – they are well defined and progress this field. The authors have collected and make use of an extensive dataset to investigate whether surface grain size can be mapped using point cloud data from SfM photogrammetry. It is great to see that 12 different rivers were used to calibrate the roughness metric. The development of this metric is rigorously undertaken and statistically informed. However, the paper isn’t just confined to a technical or methodological contribution, it also makes use of the proxies to interpret grain size patterns on bars, which lead to the observation of trends (e.g. L448-449l; L455-6), which is nice. Overall, this manuscript contributes to wider efforts to create digital models of riverscapes and to enable analysis of spatially continuous data. I agree with the observation on L585 that the findings may enable existing models of grain-size sorting to be verified or adjusted – I’d encourage pursuit of this! I recommend publication, subject to addressing the comments below.

Specific comments:

1. Thorough check for consistency in how you describe the different study areas. L159 they are called “study reaches” and L163 you say “sites”. If different phrases are needed for different things then clearly define these.
2. What is the bias that is introduced to “real world” (i.e. to the bar scale) application of the techniques when calibration plots were chose to be flat and (especially) homogenous.
3. Can you explicitly clarify whether you map grain size on just dry surfaces or whether you can also apply this to wet surfaces; can you discuss this in the discussion? I note that L548 mentions this but it’s not clear in the methods/results.
4. Is a figure to show temporal trends in grain size (2020 to 2024) missing – it is mentioned on L563?

Technical comments:

L16 Abstract : Can you be more specific about what a “proxy” is for a potentially general abstract reader?

L57 Insert “the” before Feshie

L98 Insert “sediment on” before gravel

L100 Insert “an” before approach

L178 Were shadows avoided when acquiring images? Overcast conditions? If there were shadows, what are the implications in point cloud reconstruction?

L223 You may be interested in Stott et al (2020) who, in one of their experiments, assessed the accuracy of using this DEM with the same field set up as you https://doi.org/10.3390/drones4030055

Table 2 It would be useful to present the range of the point cloud density for each site. Point cloud density can be uneven, so an average isn’t always very useful.

L300 Change “samplings” to “samples”

L303 What does “quite stable” mean? Can you rephrase/clarify

Figure 9 Is the grain size shown on the legend D50?

Section 4.1 Reid et al (2019) – already cited – developed a relationship between roughness and D50 that included bars with coarse grains (D50 > 200 mm), if you need further references on investigations that have also included large grains in the calibration.