Downscaling of MODIS NDVI by Using a Convolutional Neural Network-Based Model with Higher Resolution SAR Data
Round 1
Reviewer 1 Report
In this manuscript, the Authors propose a deep-learning based method to downscale the spatial resolution for the well-known vegetation index (NDVI), from Sentinel-2 data.
The paper is well-written in general, although some "typos" demand attention (see below). The problem is well-stated and introduced and results -are extensive- and well support the work done. References allow readers to comprehend the paper.
This reviewer liked this paper and especially the discussion section.
Some "typos",
line 204, "omitted"--> "emitted",
table 3, table 5.. not table 4!!
Table 9: use the horizontal layout.
Subsection 5.2, "Effective Ccondition for Application"--> revise
line 507, "the application One is the"--> revise,
There are more: revise them all.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The manuscript by Nomura and Oki details downscaled NDVI based on Sentinel SAR and NDVI and MODIS NDVI. A statistical / CNN based approach is used. The resulting downscaled 10 m resolution data product is relevant for the scale of individual crop rows and could therefore benefit agricultural management. The methodological structure of the paper is good, building from test cases to the intended product of downscaled MODIS.
Main comments
1a. Figures: there is no Figure 2, and no Table 7 or 8, so the numbering needs to be updated.
Also, Figures 4 through 12 are blurry / not crisp. Are they being generated with at least 300 dpi? This will enhance the impact of the results and also is important because the goal is to show high-resolution NDVI features.
1b. The Tables are likewise blurry, and in particular Table 9 is unreadable since it has a small font size. It would be better for the information to be a table natively embedded in the Document rather than as an image.
- The Introduction is interesting, but I think lacking a little bit specifically about past downscaling with Sentinel SAR (specifically on Line 91). Papers about this for MODIS land surface temperature (LST) come to mind, as does downscaling SMAP soil moisture. I also am confused by the comment about Landsat on Line 80. NDVI downsampling with Landsat has been done with 30 m resolution. It may be that the effective resolution is coarser, but that would also be true when downsampling with SAR.
- Throughout the manuscript, the word ‘seem’ is used excessively. While it can make sense when making an inference or hypothesis, typically it is used here for something that is simply describing a figure or statistical result. These instances should be made more quantitative or otherwise objective.
Also see:
L42 suggest “low use cost” instead of “low cost”
L115 “field area” => “crop field areas” ?
L122 about the double cropping, how many days in between plantings typically (which is relevant to the issue of temporal resolution)?
Ref 35 is not a proper citation to the MODIS product
Section 3 – what software package is used for the CNN model?
Figure 3: “Proposed CNN architecture” … for what? The caption should be more informative
Suggest mentioning at first use that VV and VH refer to polarization
L196 “solve mathematically” => “solved exactly” ?
L241 “2012” => “2020”
L272 “10.5” => “10/5”
L270-272 I don’t understand the “interval days” here. Is this referring the interval between available imagery from 2019 and the equivalent day in 2020?
L349 “Sertinel-2” => “Sentinel-2”
L471 “snowFigure . all” => check typo
L504 “Ccondition” => “Condition”
L507 “application One” => “application. One”
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
In the revised manuscript, the higher figure and table resolution is big improvement, among other changes. I noticed the authors also separately added a new Figure 2 and made modifications/improvements to Figures 11 and 12.
On re-reading the manuscript, there are however still some issues requiring revision before publication.
My main comment is I am not convinced of the claim discussed about achieving high temporal resolution. This is primed starting with:
“However, there is generally a trade-off between temporal and spatial resolution when using satellite images. To overcome this problem, …”
And the last sentence of the abstract concluding that “The proposed method… has the potential to realize both high spatial and temporal observation.”
It needs to be clearer that there is definitely still a tradeoff with temporal and spatial resolution (and accuracy). This method increases the number of 10 m resolution NDVI days relative to using Sentinel-2 only, but it is still a lower number than from MODIS at coarser resolution. Since it relies on cloud-free MODIS and/or Sentinel-2, it is especially limited in resolution in the growing season. This aspect/conclusion needs to be clearer, especially in the abstract.
Additional
Throughout, suggest changing ‘Modis’ to ‘MODIS’ per the official name of the instrument and wide use elsewhere.
The introduction was improved by mentioning past use of Sentinel-1 SAR and/or Sentinel-2 multispectral imaging for downscaling. However, I also believe the paper of Mazza et al. 2018 that is cited later is also relevant, since the core NDVI methodology is quite similar. This manuscript clearly extends that work by including coarse input instead of using SAR only, as well as showing a more complete application example for an area in Japan
Figure 2 y-axis label : suspect you mean ‘Fraction of’ not ‘% of’
L157 I disagree about not needing to cite the relevant MODIS dataset even if there are past examples of literature where people did not. Citing the data serves multiple purpose: it documents which processing version was used specifically in the work, points the reader immediately to where more information can be obtained since datasets involve nontrivial assumptions and calibrations/corrections, and gives credit to the relevant PI/team
How was the up-sampled 250 m Sentinel-2 NDVI calculated? [later on L470 it is mentioned ‘correctly up-sampled’, so one wonders what that involves]
Figure 4 caption : ‘(Forth)’ => ‘(fourth)’
L218 The equation given for the L1 loss function is not particularly insightful for the reader who is trying to recall the difference between L0, L1, etc. Writing as a summation would more readily the meaning
Table 2 is repeated multiple times in the document – this should be fixed
Table 5 caption: accuracy is relative to Sentinel-2 10 m, correct? Ideally, a figure/table with its caption will be relatively stand alone for the reader
Figure 9 and its caption is confusing
Table 6 – For Sentinel-2, are satellites 2A and 2B used? Likewise for Sentinel-1? I ask because the two MODIS satellites are broken out, but the individual Sentinel ones are not mentioned anywhere
Figure 11 caption: “Map in upper-left shows view of the Tsumagoi area” – presumably the view is true color from Sentinel-2? It should be noted
L452 about insufficient spatial resolution – is it the case that ‘A’ and ‘B’ on Figure 12 are sometimes in the same MODIS NDVI pixel, and sometimes in different ones? That information will help interpret the plot Figure 12(a)
The reference section is not consistent with MDPI style, https://www.mdpi.com/authors/references
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
