General comments:

This study by Mueting and Bookhagen is part of a logical sequence of studies (Kaab et al., 2017; Feng et al., 2020; Mazzanti et al., 2020; Aati et al., 2020; Aati et al., 2022...) on the processing of Cubesat Planetscope high-frequency data, and in particular on the improvement of orthorectification errors for the measurement of ground displacements, applied here to landslides. This is a very important subject, given the usefulness of these data, particularly for monitoring landslides (Bradley et al. 2019; Mazzanti et al. 2020; Dille et al., 2021; Amici et al., 2022; Lacroix et al., 2023,....).

The work carried out here is in-depth and deserves to be published. However, in its current state, the article suffers from several limitations, the main ones being: (1) the flow of the article is not always easy to follow and therefore deserves to be reorganised and rewritten partly, (2) there is a lack of quantitative data to validate the different approaches proposed, (3) the discussion on the comparison of the different approaches proposed is lacking, and shows the need to move towards a more systematic quantification of errors.

I think it should be possible to publish this study in Esurf once these major comments and all the other minor points have been taken into account.

Pascal Lacroix

Major comments

1) The reading flow is not easy to follow, due to (a) long descriptions of methods, which could be greatly shortened (in addition, a general scheme at the beginning of section 4 would certainly be very useful to explain your processing chain from L3B or L1B images), (b) part of the description of the results included in the figure legends, (c) (too) many descriptive figures (I think a better selection of figures should be made. For example, figures 5 do not add much to understanding and can be placed in the supplementary material. Figures 4 and 6 illustrate the same effect), (d) the method section also includes results.

2) The validation of the results is not really quantitative. Statistics should at least be provided to show the improvement of the different steps, for all sets of images. Here, a single histogram is shown for a correlation between 2 images (Figure 14). Why not extract a statistic (SD for example) from this histogram and compare it for the different steps for all pairs?

In addition, you could provide more quantitative validation by comparing your results with field measurements, which exist at least on the Siguas landslide (Lacroix et al., 2019).

3) The authors use 2 different processing approaches to extract displacement fields, using L3B or L1B images. I think there is a lack of clear discussion on which of these 2 approaches is more efficient. This discussion should be based on a more quantitative assessment of the errors on each of the processed pairs (see my previous comment). For me, this discussion should also include a systematic analysis (for all pairs) of subframe misalignment errors. The authors claim that they are reduced in the latest acquisitions. Could the author clarify why they have come to this conclusion, and when this improvement was made?

Detailed comments

L15: "geoscientific": I would rather say "geomorphic".

L36-37: "landslides are prone to orthorectification errors": It would be useful to quantify this orthorectification error. I suggest reviewing all the uncertainties associated with the use of PlanetScope data for landslide studies (Bradley et al. 2019; Mazzanti et al. 2020; Dille et al., 2021; Amici et al., 2022; Lacroix et al., 2023, ...).

L75: It would be interesting if you also mentioned that monitoring already exists on the Siguas landslides, which could be used to validate your results (Lacroix et al., 2019). See my main comment no. 1.

2.1: Are there independent estimates of the speed of the Del Medio landslide?

Figure 1: As things stand, the black and blue lines mentioned in the legend are difficult to see. Is there any real point in showing the road network?

L110: "NIR measurements are stored at the green pixels of the RGB Bayer-mask (Planet, 2022a). "This is not clear. Besides, is there any point in knowing this information? In general, I think authors should simplify their text to make it easier to read (see my main comment No. 1).

L114: "NIR band is captured at a different time": Can you specify what the timeframe is?

Figure 3: This is a nice figure to show the different errors. I would simply reverse the order of the legend so that it corresponds to the order of the sub-figures: (1) DEM error (A), (2) striping errors (B, C, D), (3) overall shifts between scenes (C, A), (4) stereoscopic errors (D).

L155-156: It should be noted that the error associated with a global offset is classically corrected for slow slide studies using PlanetScope, which significantly reduces the errors (see also my comment on the uncertainty associated with PlanetScope images of slow slides l36-37).

L240: L1B images are also available in clipped format.

Figure 5 : « Scenes acquired from an opposite view direction at high view angles are

strongest affected by orthorectification errors. » : Opposite view direction to what ? why should orthorectification errors be stronger with some specific viewing angles ? This sentence is not clear.

Furthermore, Figure 5 may not be necessary to understand the study. In fact, it is mentioned only once in the text. Could you place this figure in the supplements?

Lines 250-287: This section can really be reduced. Figure 4 illustrates this well. I also wonder if this section should not be mixed with section 3.1, when the effect of orthorectification errors is illustrated in figure 2. In this section, you do not propose a method for reducing this error, but you do illustrate it. In my opinion, it should not be included in the "data and methods section".

Reorganisation: Lines 288 to 313: I get the impression that this section is a bit vague and that the flow is not easy to follow because it's a mixture of methods and results. If I understand you correctly, you identify the acquisition parameter that allows you to form pairs and reduce uncertainties while correlating them. I have the impression that you could state this much more clearly and separate the methods from the application to the data. The choice of figures also makes things less easy to follow: Figure 6 is closely related to Figure 4 in terms of illustrating the problem (perhaps one of the two figures could be placed in the supplementary material?) Figure 7 is an application of the methods that shows the important effect of the actual azimuth of observation. Figure 8 shows your results once the groups have been created.

In the same section, it is not clear, once your pairs are created, how you will use them to create a time series of movements. In fact, there is no possible relationship between the different groups of images. How do you put them back together? I have the impression that a general diagram at the beginning of section 4 would be useful to explain your processing chain based on images L3B or L1B.

Line 375: I have the impression that the method you describe has already been described and used by Berthier et al (2007). You can certainly simplify your text by refering to it.

Line 394: MPIC-OPT is not strictly a correlator but a processing chain that does more than correlate. The correlator behind MPIC-OPT is Mic-Mac (Rupnik et al., 2017).

Line 399: Why do you use 35x35 pixel windows? Did you do several trials before choosing?

Line 399: I would also delete "slightly larger correlation".

Figure 11: There is no reference to sub-graphs C and D.

Line 417: You mention striping effect due to the misalignment of the subplots, but I assume that your polynomial (line 420) does not correct this effect. How effective is your polynomia at correcting other artefacts when you have such effects?

Line 435: The speed of the landslide may vary over time, but you are assuming here that the speed is constant over the period in question. This needs to be made clear. In fact, you mention this transient in the caption to Figure 12. This assumption could be verified by comparing satellite measurements with field data. These validation data are available on the Siguas landslide (Lacroix et al., 2019). See also my following comment on the results validation.

By the way, your Figure 13 seems to show that the standard deviation of velocity is significantly reduced when using manually orthorectified 1B products compared to correlating 3B products from the same group. To me, this means that the higher standard deviation observed on the landslide in Figure 12 does not come from transient motion but rather from orthorectification errors in the 3B products, even if you select the pairs. Can you comment on this point?

Line 436: remove an "only".

Figure 15: I don't quite understand how you obtain these time series. I understand that you correlate either the L1B data from the same group or the L3B data from group 4, but do you correlate them all within each group or do you only correlate those that are separated by the shortest time to see the transients?

Line 541 : The correct reference is Lacroix et al., 2019 not 2015

Line 550-553 : Are you removing the low quality pixels from the « good pixel map » ? In this case it should highly remove the changes in soil occupation, and therefore the errors. Are you not sure that the higher standard deviation with time in the Siguas case study is not caused by the motion of the landslide that occupies a quite important area of the image ?

Rather than a hypothetic section on the « Transferability to other regions and targets », I would have rather see a discussion on which of the L1B or L3B processing should we use (See my major comment n°3). From the Figures you show, manually orthorectified L1B sounds more efficient, except for the sub-frames alignment. However it lacks this analysis for all the scenes processed. Furthermore, is the sub-frame alignment really better now ?

Figure 15 and lines 610-612 : Can you explain how you obtain the shown uncertainties and add uncertainties on your velocity estimations ?

Dear authors,

Please find my comments in the attached pdf.