A Scale-Separating Framework for Fusing Satellite Land Surface Temperature Products
Round 1
Reviewer 1 Report
The authors of the manuscript entitled "A Scale-Separating Framework for Fusing Satellite Land Surface Temperature Products" tried to generate high resolution LST data by developing scale-separating framework. The developed framework provides improved error propagation against bilateral filtering which is a novel approach. The manuscript is well written and is simple to understand to the readers. However, there are minor English corrections that the authors need to do prior to publication. I recommend authors to modify the methodological framework a bit self explanatory. For example; Downscale and upscale of the images are being done but it is not understandable from the work flow. Its better to design the workflow so that it can explain by itself.
Author Response
Thank you for your review! We have uploaded the responses and the updated manuscript. Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The manuscript presents a new and very interesting methodology. It is clearly organized for a specialized audience and reports a very useful combining technique.
The introduction is very complete and informative.
The methodology is very well organized reporting in a clear way all the various steps needed to achieve the results.
The discussion is clearly presented and convincing.
I suggest the publication in the present form.
Author Response
We appreciate your careful review and your recognition! We synthesized the suggestions from all the reviewers and have updated the manuscript.
Reviewer 3 Report
This paper presents the difference between LST and air temperature, by which air temperature is also affected by the condition of wind speeds.
It is creative to develop the comparison of air temperature with LST through the urban area and rural area by local route of air temperature measurements. However, the data validation of LST calculated values with measurements. It may also need to indicate what is the relationship between LST and air temperature or if they have any relationship.
So it needs a bit more discussion about this matter or if it needs to be validated. In addition, some more references should be updated or added.
In short, it needs a minor revision.
Author Response
Thank you for your review! We have uploaded the responses and the updated manuscript. Please see the attachment.
Author Response File: Author Response.docx
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
Generally, it is a very interesting manuscript focusing on thermally sharpening the MODIS LST products. My concerns are listed as follows,
- In section 2.1, basic information of this study area, including the geographical domain, natural conditions, and human activities, should be elaborated, given that not all the readers know this area well.
- In section 2.2, the authors should clearly address how to retrieve the LST from Landsat-8 OLI/TIRS data, rather than simply citing the references. How to correct the land surface emissivity and which TIRS band were used？
- Since MODIS and Landsat derived LST products have the different spatial resolutions. Then how to perform the fitted regression (e.g., Figure 3a) from each pair of these two LSTs must be detailed. Besides, for Figure 3b and 3c, how about the sample size?
- The first researching goal of this study is “To address the non-linearity of inter-sensor LST relationships with incorporation of a neural network,”,but which figure shows such non-linearity?
- I do not think Figure 14d shows the statistically positive regression between sharpened MODIS LST and air temperature. How can such regression be validated with bicycle measurement of air temperature on one date? Besides, the air temperature varied along the routes across different landscapes.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 2 Report
The present work the authors developed a fusion framework to generate new LST data with fine spatiotemporal resolution. The framework was experimented with MODIS and Landsat LST data. Differetnt method were showed (sensor to sensor biases, linear stretching across time, neural network) and a accurate analysis of training and validation Loss and Accuracy assessment was made. The error analysis showed very googd results in terms of RMSE and the comparison between the sharpened LST and the air temperature measured with bicycle-mounted mobile sensors revealed the roles of impervious surface fraction and wind speed in controlling the surface-to-air temperature gradient in an urban landscape.
I believe that this work adds something innovative in the studies of urban environments in terms of heat island as MODIS can give us more frequent data over time but it can keep the Landsat resolution at spatial scale which is fundamental for an urban scale.
Author Response
We appreciate your careful review of the paper! We have revised the manuscript according to the suggestions from all the reviewers and have uploaded it for you to check.
Reviewer 3 Report
This paper proposes a method based on the concept of scale separation to construct LST databases characterized by high temporal sampling and high spatial resolution.
The subject of the work is very popular. There are many papers, published during the last decades, which discuss the joint use of Landsat and MODIS data for constructing multiresolution LST databases. The goals of the work are well explained, but the methodological part is not nearly so clear and some a priori hypotheses are questionable.
The main basic assumption is that spatial details at the sub-kilometre account for change on a seasonal basis but only their 1km mean varies on interannual time scales.
The authors affirm that “The validity of this assumption is supported by the 1:1 comparison of the within-patch variations in LST between the target and the reference Landsat image (Figure 3c)” (lines 281-283). Just the careful inspection of this figure should instead suggest great caution. The distribution of the residuals around the grey line is very large. In particular, pixels with values of LST near to zero in one of the two Landsat images correspond to values ranging from -15 to 15 Ko in the other one. In many cases, high positive values in one of the images correspond to large negative values in the other one. In brief, the proposed method generates many unrealistic data.
In my opinion, the work is based on a premise that has not been sufficiently justified and, on the contrary, appears to be very questionable. The spatial heterogeneity of LST is mostly the consequence of the heterogeneity of the surface emissivity, generally synthesized by the NDVI, whose values change at field and canopy level both on intra- and inter-annual scales.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
The article is interesting and well prepared and should certainly be published. Minor remarks regarding additions and corrections have been placed next to the text of the article in the PDF file.
My general conclusion is that the proposed methodology is very complex and it will be difficult to apply it in practice without the appropriate software (computer algorithm). Moreover, the use of the IMP index limits the usefulness of the procedure to the USA. Do the authors have any proposal of an alternative to the IMP enabling a wider geographic application of the described methodology of time-space sharpening of LST data? Is the significance of the IMP really big?
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
1) Please add the p values into Figure 15 to show the statistical significance.
2) The authors' contribution to this researching domain should be highlighted. Besides, the limitations of this study and future researching plans for improving the quality of such studies should also be addressed.
Reviewer 3 Report
I understand the authors' reasons, but I still do not share their confidence in the reliability of the proposed approach.
- They affirm that spatial details at the Landsat resolution only account for seasonal variability. However, if data recorded in different DOYs (say DOY1 and DOY2) do not change from one year to another one, also their difference DOY2-DOY1 does not change. In practice, also the seasonal curves at high resolution do not change. In such a scenario, only the MODIS scales are actively involved in the LST dynamics. Therefore, high-resolution LST deviations from the 1-km average do not describe any relevant variability. In my opinion, this assumption is very strong and hypothesized without any scientific justification.
- Due to the strong correlation between LST and NDVI, the reasoning above also applies to vegetation phenological curves that show instead evident interannual variability and large spatial differences at local scale. Individual and environmental factors (weather conditions at the micro- and macro-scale, soil conditions, water supply, diseases, competition, etc.) influence the annual plant phenology during each specific year but in the scenario proposed in this paper, interannual differences emerge only at coarse resolution (!).
More in general, the author’s reply did not help me to change opinion. An RMSE value estimated for a large image is an average “global” measure. We do not know if very large local errors occur randomly over the area or they come from peculiar patch configurations (e.g., transition zones). Users interested in high resolution studies of possibly critical patches could obtain misleading results. In addition, we do not know if the RMSE value depends on the season (why not?) or on the meteorological conditions characterizing the target and reference years (are they similar?). In brief, the low RMSE reported in figure 3 is not conclusive.
As for the question of alternative methods that use the closest available information, this approach is physically well founded due to the continuity of temperature in time. This is a common practice also to reconstruct data missing in air temperature data from ground stations. It is evident that when the closest value is very far, possibly in a different season of the year, the reconstruction cannot work well, but that is not the point.
No real scientific evidence of the rationale underlying the proposed method has been provided. I think that this is a crucial point that needs in-depth analyses before proposing a dataset as a reliable tool for applicative purposes.
