the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 26 May 2026
A pan-tropical 5-km monthly L-band vegetation optical depth dataset from pan-sharpening-based downscaling
Abstract. L-band Vegetation optical depth (L-VOD), as a microwave-derived vegetation indicator, has been widely applied in the monitoring of vegetation dynamics. However, the spatial resolution of 25-km or coarser in existing L-band VOD products limits their applications in ecological monitoring requiring a higher level of spatial details. To mitigate this limitation, we introduce a pan-sharpening-based downscaling method to improve the spatial resolution of L-VOD. By fusing the spatial structural features of the aggregated 5-km resolution European Space Agency Climate Change Initiative (ESA CCI) aboveground biomass (AGB) product, the SMOS L-VOD product over tropical regions was downscaled to generate a monthly 5-km resolution L-VOD dataset spanning 2015 to 2021. The downscaling model demonstrated high accuracy, with a correlation coefficient (R2) of 0.95 and a root mean square error (RMSE) of 0.11 when comparing the simulated 25-km L-VOD (L-VOD25kmsim) with the original L-VOD (L-VOD25km) product. Spatially, the 5-km resolution L-VOD (L-VOD5km) yielded a strong correlation with above-ground biomass (R=0.91, R2=0.86), and temporally dynamics, it accurately characterized the LAI variations of short vegetation and forest area loss at the pixel level over the study period. The results demonstrate that our downscaling method can effectively enhance the spatial resolution of L-VOD while preserving its original spatiotemporal dynamics, and is capable of capturing forest disturbance. This dataset can be downloaded at https://doi.org/10.11888/RemoteSen.tpdc.303391 (Shi and Fan, 2026).
- Preprint
(4989 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 26 Jul 2026)
-
RC1: 'Comment on essd-2026-193', Anonymous Referee #1, 26 Jun 2026
reply
-
RC2: 'Reply on RC1', Anonymous Referee #2, 23 Jul 2026
reply
The paper is correctly written while the English could probably be improved as some sentences are not crystal clear.
The content seems sound to me and the questions raised on this type of publication are probably more generic and not always specific to the manuscript itself.
General comments
It s more semantic than anything else but I feel that it is not correct to speak about spatial resolution as done in this document. Spatial resolution is linked to the sensor characteristics and thus the the area from which the measured energy comes from. It is not to be mistaken with the sampling. In terms of information content etc this is important. I would thus recommend to replace all the “spatial resolution” by "spatial sampling" in this manuscript.
The second general comment is more on the approach itself. If it is a good this to try to have higher resolution L-VOD maps one may question (as the authors do incidentally) to what the product will be compared the outcome. The authors go to some details on the four pronged approach, but unfortunately it does not really address the issue. The history is full of approaches/ models being validated against dubious (but only available for sure) data sets to be proven wrong (resp. true) while validated (resp. said wrong).
The first deals with usual metrics. Comparing the two LVOD -25km is straightforward but only proves that the method does not go completely awry, it is simply a sanity check and does not prove much. The fact that it works well over dense forest does not prove anything as it is where you would expect the least variability.
For the temporal and spatial assessment, I must have missed a step as, to my understanding, the output is compared to the AGB map which was used to elaborate the dis-aggregated product. Maybe the authors could detail this a bit further.
The comparison with LAI also does not bring much. First LAI and VOD are two different things. They do nand ifot look at the same characteristics of the vegetation (chlorophyll vs. water content) not the same media (top surface for LAI vs. more volume for L band). So no wonder the R are so low (and sometimes negative) except for dense forests where both tend to saturate.
The forest loss is probably the most interesting “validation attempt but again it is not very satisfactory. Comparing a curve to a point (fig 10) is interesting but does not drive any nail in unfortunately. The explanation for the difference in NDVI vs LVOD recovery is not very convincing. For me first again we are not comparing the same things so it does not prove much other than that L-VOD (Water and thus BIOMASS in these regions) do recover slowly (what we expect) but also if the area is deforested , it is to plant crop which have been selected to produce and thus tend to have higher LAI (NDVI) that natural vegetation. So for me this means that the area sees forest replace by crops.
One final general comments. The b=references are nice but seem to be a bit partial towards the authors. Common, practice for sure but it is always strange to see that studies by some authors seem to be under represented.
In conclusion a paper addressing an important point but with some flaws in the validation meaning it does not go beyond existing papers (as the methodology has already be used and published in other papers). So not really worth publishing to me.
Citation: https://doi.org/10.5194/essd-2026-193-RC2
-
RC2: 'Reply on RC1', Anonymous Referee #2, 23 Jul 2026
reply
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 143 | 61 | 11 | 215 | 14 | 11 |
- HTML: 143
- PDF: 61
- XML: 11
- Total: 215
- BibTeX: 14
- EndNote: 11
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General Comments
This manuscript proposes a new 5-km L-band Vegetation Optical Depth (L-VOD) dataset by downscaling the 25-km L-VOD product using a 5-km Aboveground Biomass (AGB) dataset through a pan-sharpening-based framework. The resulting product is comprehensively evaluated against higher-resolution auxiliary datasets, including AGB and vegetation indices, and further demonstrated through a spatiotemporal analysis of forest loss. Overall, I appreciate the authors' effort in developing this dataset and conducting extensive experiments and comparisons. The topic is relevant and the generated product has potential value for the remote sensing community. However, I believe the manuscript would benefit from improvements in the presentation, particularly in the description of the methodology and the interpretation of the results. Therefore, I recommend major revision.
Major Comments
Study area description (Section 2).
The study area section would benefit from additional background information on the ecosystems of the two selected regions to better motivate their selection and provide readers with more context. In addition, please explicitly explain what P1 and P2 represent in Figure 1.
Potential numerical instability (Section 3.2, Eq. (2)).
Since the injection gain is computed as (g = L\text{-}VOD_{5km,NNR}/AGB_{5km,LowPass}), what happens when (AGB_{5km,LowPass}) approaches zero? Would this lead to numerical instability or extremely large gain values? Please explain how such cases are handled.
Description of the iterative optimization of the injection gain (around Lines 190–205).
The description of the iterative optimization of the injection gain (g) (Eqs. (2)–(4)) is difficult to follow. Since this section describes an algorithm rather than a mathematical derivation, I strongly recommend presenting it as pseudocode or an algorithm box.
At present, it is unclear what variable is actually being optimized and how the optimization is performed. Specifically, Eq. (3) states that g is optimized by maximizing the Pearson correlation between the downscaled L-VOD and AGB, whereas the subsequent text only describes progressively reducing the upper bound (g_{limit}) and clipping g. It is therefore unclear whether g is recomputed following some math optimization rule for equation (3), or whether only the clipped version of the initial gain map is updated.
A pseudocode description would clearly specify the initialization, the variables updated in each iteration, how Eq. (1) is repeatedly evaluated, how the correlation is used during optimization, and what constitutes the final optimized gain map. This would greatly improve both the clarity and reproducibility of the proposed method.
Figure 2 (algorithm flowchart).
The flowchart does not sufficiently clarify the iterative optimization process. In particular, the step "g = adjusted g" is ambiguous, as it is unclear how the adjusted gain is obtained. Similarly, the role of (\Delta R) in the optimization is not well illustrated. Revising the flowchart to explicitly show the update rule for g (or (g_{limit})) would substantially improve readability.
Comparison with existing downscaling methods.
The manuscript evaluates the proposed product against auxiliary datasets but does not compare the proposed downscaling approach with existing methods. Since the Introduction discusses alternative approaches (e.g., SAR-based methods, statistical downscaling, MF-MRA), it would strengthen the manuscript to compare the proposed method with one or more representative baselines and demonstrate its advantages.
Forest loss analysis.
The forest-loss analysis could be further strengthened. Since the Hansen Global Forest Change product has well-known limitations (e.g., it represents tree-cover loss rather than a broader definition of forest loss, with trees defined as taller than 5 m), these limitations should be briefly discussed in either the Results or Discussion section.
In addition, I suggest performing the analysis at the regional level rather than emphasizing pixel-level dynamics (Figure 10c,d), which may be sensitive to uncertainties in the Hansen product. Furthermore, Figure 10 only shows a single blue bar for forest loss in 2015. It would be more informative to visualize annual forest loss throughout the entire study period (2015–2021).
Minor Comments
Section 3.3 (Accuracy assessment).
Consider formatting the subsection headings such as "(i) Evaluation of the downscaling method" in bold to improve readability.
Figure 4 (d, f).
Why is the standard deviation (error bars) of the downscaled (L)-VOD({5km}) consistently smaller than that of the original (L)-VOD({25km})? Since the downscaled product is expected to capture more spatial details, one might expect increased variability. Please provide an explanation.
Figure 5.
Consider including the corresponding AGB image and an optical Sentinel-2 image alongside the L-VOD visualization to facilitate interpretation.
Figure 6 (j–l).
Since all three panels show relationships involving (L)-VOD(_{5km}), I suggest using a consistent x-axis label across all three plots.
Section 4.2 (Accuracy assessment).
The manuscript states that the performance was evaluated using data from 2019. Why was only a single year selected instead of evaluating the entire study period (2015–2021)?
Figure 8.
Consider adding labels such as "DRC" and "Brazil" directly to the figure to improve readability.
Consistency between Figures 5 and 8.
Figures 5 and 8 appear to use the same study regions. To improve consistency, I recommend using the same naming convention throughout the manuscript. For example, instead of using Region 1/Region 2 in Figure 5 and Rondônia/DRC in Figure 8, consistently use Brazil and DRC.
Terminology consistency (around Line 365).
The manuscript inconsistently uses AGB(_{100}), AGB(_{100m}). besides, based on the methodology, it seems more appropriate to refer to AGB(_{5km}), since the downscaling uses the aggregated 5-km AGB product rather than the original 100-m spatial details.
Section title (Section 5.2).
The current title "Uncertainties and Perspectives" would be more appropriately renamed "Limitations and Perspectives", as the section primarily discusses methodological limitations and future research directions.