the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 13 Mar 2023
HSR-GEE: A 1-m GEE Automated Land Surface Temperature Downscaling System over CONUS
Abstract. Downscaling remote sensing-based Land Surface Temperature (LST) datasets is of paramount importance for multiple research fields including urban heating, irrigation scheduling, volcanic monitoring, to name a few. Even with the constant development in technology, especially improvements in spatial, temporal and spectral resolutions of satellite sensors, global satellites are limited to a 60-m LST datasets at best. In this study, we use the massive computation power and large dataset directory found in the Google Earth Engine (GEE) platform to design a 1-m fully-automated, open-source and user-friendly LST downscaling system, named High Spatial Resolution-GEE or HSR-GEE. It has the ability to downscale Landsat-8 LST, combined with the National Agriculture Imagery Program (NAIP) images, into 1-m HSR LSTs over the CONUS and at Landsat-8 overpass time. Using only red, green, blue and near infrared bands, HSR-GEE implements multiple machine learning approaches, including, Robust Least Square (RLS), Random Forests (RF), and Support Vector Machine (SVM), along with comparison to two commonly-known and classical methods: the disaggregation procedure for radiometric surface temperature (DisTrad) and thermal sharpening (TsHARP). We validate HSR-GEE outputs against multiple airborne thermal images over the USA. We obtained a MAE of 1.92 °C, 2.53 °C, 1.33 °C, 3.42 °C and 3.4 °C for the RLS, RF, SVM, DisTrad and TsHARP, respectively. With RF showing visually salt and pepper effect and SVM a Land Cover/Use form, the RLS appears to be most suited for 1-m LST downscaling. HSR-GEE is proposed as a high-potential system aiding researchers from different backgrounds to advance their research. HSR-GEE remains the only available 1-m GEE-based downscaling system that is able to derive the needed high resolution LST information in a matter of seconds and in five different approaches (i.e., RLS, RF, SVM, DisTrad and TsHARP). The research community is invited to implement this dynamic system over CONUS and enhance it if deemed necessary.
- Preprint
(1174 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
CEC1: 'Comment on egusphere-2023-11', Juan Antonio Añel, 06 Apr 2023
Dear authors,
Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".
https://www.geoscientific-model-development.net/policies/code_and_data_policy.htmlThis is especially unfortunate, as the Topical Editor of your manuscript has already requested twice you move the code to one of the suitable repositories, as we can not accept either the storage in Google Cloud or that it is necessary to log in to get access to it.
Currently, the situation with your manuscript is irregular, as it should not have been accepted for Discussions as it does not comply with our policy. Therefore we will have to reject your manuscript for publication unless you solve this situation in a prompt manner.
In this way, please, publish your code in one of the appropriate repositories, and reply to this comment with the relevant information (link and DOI) as soon as possible. Also, if the topical editor requests you a reviewed version of your manuscript, you must include in it the modified 'Code and Data Availability' section, including the DOI of the code.
I must emphasize again that failing to comply with this request will result in the rejection of your manuscript for publication.
Juan A. Añel
Geosci. Model Dev.
Citation: https://doi.org/10.5194/egusphere-2023-11-CEC1 -
CC1: 'Reply on CEC1', Mario Mhawej, 13 Apr 2023
Dear Dr. Añel,
Many thanks for your comment!
Kindly note that indeed the Topical Editor has previously requested to change the code and we were able to reply to him and explain the situation. As we completely understand the "code and data policy" of the journal, it is important to note that the issue lies in the Google Earth Engine (GEE) and it is not from our end. More precisely, HSR-GEE was developed based on the Java Script language which is fully built on the GEE platform to save on time and resources for users, as it uses the massive computation power of GEE along its very large databases. The only drawback is that GEE requests to have a google-based username to run the code (the source can be accessed freely but cannot run without credentials). Still, even as authors and developers of the HSR-GEE we do not have any access whatsoever to who has used the provided link or which credentials were used to access the code. As a result, reviewers anonymity is reserved.
Anyhow, we have tried to build an application from the HSR-GEE to overcome the fact that users need an account the run the code, which has required a lot of resources and time, only to adhere to the requested policy as suggested. Now, anyone can run the HSR-GEE application from: https://mariomhawej.users.earthengine.app/view/hsr-gee
It is important to note, however, that this application has its own limitations, related to the user limitation capacity at the GEE, which has added few constraints on the full usage of the HSR-GEE system. This include:
- Limited study area at each run;
- Longer processing time;
- Capacity limit issue.
We will be contacting the GEE team to inquire more about how to overcome these issues. In the meantime, we believe reviewers and any interested party can access and run freely the HSR-GEE application (https://mariomhawej.users.earthengine.app/view/hsr-gee) or via credentials to access and run the code wihtout the application as mentioned in the manuscript. Both source codes can be accessed on https://github.com/mariomhawej/HSR-GEE/
We hope this helps,
Kind regards,
Mario M.
Citation: https://doi.org/10.5194/egusphere-2023-11-CC1 -
CEC2: 'Reply on CC1', Juan Antonio Añel, 13 Apr 2023
Dear authors,
I understand your explanation; however, it was your choice to use the GEE system for your work, and despite choosing a system that is not open, you are trying to publish your work in a journal with a strict policy on access to the code.
Also, I think that you have not understood our request adequately. It is not that we need to be able to run the code but that we need it stored in an open repository. You point to one in GitHub; however, again, our policy is very clear about the fact that GitHub is not acceptable for scientific publishing. I would ask you to read the code and data policies of the journal and comply with them, as you want that your work is considered for publication in Geosc. Model Dev.
GitHub itself instructs authors to use other alternatives for long-term archival and publishing, such as Zenodo (GitHub provides a direct way to copy your project to a Zenodo repository). Therefore, please, publish your code in one of the appropriate repositories.
Also, in the GitHub repository, there is no license listed. If you do not include a license, the code continues to be your property and can not be used by others. Therefore, when uploading the code to the repository, you could want to choose a free software/open-source (FLOSS) license. We recommend the GPLv3. You only need to include the file 'https://www.gnu.org/licenses/gpl-3.0.txt' as LICENSE.txt with your code. Also, you can choose other options: GPLv2, Apache License, MIT License, etc.
Please, reply as soon as possible to this comment with the link to the new repository (and DOI) so that it is available for the peer-review process, as it should be.
I insist that if you continue failing to comply with the requirements of the journal, your manuscript will be rejected.
Juan A. Añel
Geosci. Model Dev. Executive Editor
Citation: https://doi.org/10.5194/egusphere-2023-11-CEC2 -
CC2: 'Reply on CEC2', Mario Mhawej, 14 Apr 2023
Dear Dr. Añel,
Many thanks for understanding and for your continued support!
Here is the script DOI: https://doi.org/10.5281/zenodo.7828718
Many thanks,
Kind regards,
Mario M.Citation: https://doi.org/10.5194/egusphere-2023-11-CC2
-
CC2: 'Reply on CEC2', Mario Mhawej, 14 Apr 2023
-
CC1: 'Reply on CEC1', Mario Mhawej, 13 Apr 2023
-
RC1: 'Comment on egusphere-2023-11', Anonymous Referee #1, 28 May 2023
This paper aimed to develop a method to generate 1 m resolution Land Surface Temperature (LST) data from Landsat LST images. They build a relationship between 900-m resolution LST (upscaling for original Landsat LST band) with 30-m resolution Landsat visible bands and NIR band. Then, they downscaled the Landsat original LST band to 1 m by transferring this relationship via RLS, RF, and SVM methods and the 1 m NAIP visible and NIR bands. They also compared regression and machine learning (RLS, RF, and SVM) methods with traditional methods (DisTrad and TsHARP). Although, the regression method, especially the RLS showed the best accuracy when compared with airborne LST data in three validation areas. I still have some concerns before this manuscript is recommended for publication.
Major comments:
- The LST bands in GEE are resampled to 30 m, in fact, the original resolution of Thermal infrared 1 bands for Landsat 8 TIR is 100 m, when you used 900 m LST as an input variable for the regression and machine learning, the scale is 9 not 30.
- The principle between object reflection and emission is different. In the visible bands, objects reflect different amounts of light depending on their color and composition. In the NIR bands, they absorb or reflect light based on their structural and moisture content. While the emission ability is dependent on the thermal properties of objects. Meanwhile, The spatial heterogeneity of the temperature of ground objects is not the same as its reflectance band. Why can visible light and near-infrared bands be used for downscaling? The authors should make it clear.
- The spatial pattern of temperature in Figure 1 and Figure 5 (in Clayville) do not show high quality, the temperature change in adjacent areas is not continuous. For example, in the middle of the right part of Figure 1, the blue is an obvious sudden change in temperature, and in Figure 5, the temperature extracted with HSR RSL method in claryville (bottom figure) shows an obvious cross line. The sudden change and discontinuity in spatial patterns should be improved.
Citation: https://doi.org/10.5194/egusphere-2023-11-RC1 -
AC1: 'Reply on RC1', Mario Mhawej, 22 Jun 2023
This paper aimed to develop a method to generate 1 m resolution Land Surface Temperature (LST) data from Landsat LST images. They build a relationship between 900-m resolution LST (upscaling for original Landsat LST band) with 30-m resolution Landsat visible bands and NIR band. Then, they downscaled the Landsat original LST band to 1 m by transferring this relationship via RLS, RF, and SVM methods and the 1 m NAIP visible and NIR bands. They also compared regression and machine learning (RLS, RF, and SVM) methods with traditional methods (DisTrad and TsHARP). Although, the regression method, especially the RLS showed the best accuracy when compared with airborne LST data in three validation areas. I still have some concerns before this manuscript is recommended for publication.
Major comments:
- The LST bands in GEE are resampled to 30 m, in fact, the original resolution of Thermal infrared 1 bands for Landsat 8 TIR is 100 m, when you used 900 m LST as an input variable for the regression and machine learning, the scale is 9 not 30.
We would like to thank you for this comment. Indeed, the used USGS Landsat 8 Level 2, Collection 2, Tier 1 TIR product is resampled to 30 meters to match multispectral bands. The scale is 9 instead of 30. The manuscript was improved accordingly.
- The principle between object reflection and emission is different. In the visible bands, objects reflect different amounts of light depending on their color and composition. In the NIR bands, they absorb or reflect light based on their structural and moisture content. While the emission ability is dependent on the thermal properties of objects. Meanwhile, The spatial heterogeneity of the temperature of ground objects is not the same as its reflectance band. Why can visible light and near-infrared bands be used for downscaling? The authors should make it clear.
We would like to thank you for this interesting comment. As discussed in the Introduction section, two main approaches are used for the downscaling purposes. The first is physical-based and the second is statistical-based. Both, particularly the statistical-based approaches are widely available in the literature. These later have used different kind of variables, even the most remote or indirectly related to LST (such as population, income level, etc.). Still, it is important to note here that the produced HSR LST is based on both L8 LST and HSR RGBN images, not only HSR RGBN.
If the reviewer is suggesting to further assess bands and produce an index such the Normalized Difference Vegetation Index (NDVI) or any other index which could somehow depict a physical properties of land features, then, it is important to note the main objective of this paper. The aim of this paper is to use the freely-available 1-m NAIP multi-spectral images in the downscaling of openly available low resolution LST information. Such downscaling process would benefit multiple disciplines in science, including urban-, environmental- and agricultural-based studies, to name a few. Even when using commercial HSR images which come with limited number of bands, and where in some cases the addition of bands would generate more fees, in this study, we tried to retrieve an approach that uses only limited number of bands, namely red, green, blue and NIR, which should be available in all HSR images in different sensors, coupled with the freely available L8 LST images. Of course, including the short-wave infrared (SWIR) band, particularly for urban-based studies is always recommended, but it would generate further limitations for users to implement the HSR-GEE algorithm if the SWIR band is not available. The same goes for generating albedo, an interesting variable for LST calculation, but requiring 5 different bands in Landsat-8, very difficult to retrieve in HSR images.
To summarize, authors of this paper have first intended to include some of the most common indices or parameters, such as NDVI, LAI, or albedo, to cover structural and moisture contents but due to the availability of only four bands (i.e., RGBN bands) and because of the sustainability of the system, the HSR-GEE algorithm was proposed using a widely acceptable statistical-based approach and generating HSR LST based only on HSR RGBN bands and L8 LST values. Further studies could try to include other bands and factors whenever possible as well as including in the analysis and interpretation the potential effect of thermal anisotropy, and more importantly, differences in albedo for different surfaces. Again, thank you for pointing this matter out.
- The spatial pattern of temperature in Figure 1 and Figure 5 (in Clayville) do not show high quality, the temperature change in adjacent areas is not continuous. For example, in the middle of the right part of Figure 1, the blue is an obvious sudden change in temperature, and in Figure 5, the temperature extracted with HSR RSL method in claryville (bottom figure) shows an obvious cross line. The sudden change and discontinuity in spatial patterns should be improved.
We would like to thank you for this comment. To describe more our approach, below is an example:
When three roofs are seen from space, the RGBN values of each specific roof are retained. These values do not directly provide the LST values as it would be not accurate nor physically-based and this is due to diverse factors we all aware of. What we did instead is we took the main LST value at a higher spatial resolution, namely 30-m, and "adjust" it on a subpixel level based on the 1-m RGBN values using multiple statistical and ML based approaches. As a result, in our approach the coarser resolution LST values were used and slightly enhanced to capture the change in roof LST values between large pixels.
If a land-cover or an emissivity-based based approach is used instead, the three roofs would fall under same class and thus having the same emissivity and class value. Below a descriptive table.
Roof #1
Roof #2
Roof #3
Number of classes
Our RGB approach
30 R, 40 G, 70 B, 110 NIR
130 R, 50 G, 80 B,
21 NIR
30 R, 40 G, 69 B,
111 NIR
3
Land cover or emissivity-based approach
One value for the roof: for instance, 0.2
Same value for the roof as it is in the same land cover type: for instance, 0.2
Same value for the roof as it is in the same land cover type: for instance, 0.2
1
As a result, the newly generated 1-m HSR-GEE is based mainly on 30-m L8 LST and "tuned" with the RGBN values. Because of that the sudden change and discontinuity in spatial patterns in the 30-m L8 is transferred into the 1-m HSR product (a smoother pixels’ transition can be seen for instance in Figure 2 due to a smoother transition in the L8 LST product’s pixels). This "pixelation" is mostly due to cloud coverage and can be overcome by applying a resampling/focal/neighborhood analysis approach at the end of the process but it would erode the data and thus was not recommended in many downscaling studies. Another option would be to include a data imputation technique which is beyond this research. Still, it can be tested in our future research to enable a smoother transition between 1-m LST pixels.
It is important to note however that this did not interfere with the product quality as the generated outputs were validated following the best practices for satellite-derived LST product validation described in Guillevic et al. (2018) using airborne TIR images and yielding very promising accuracy.
We would like to thank you for taking the required time to provide us with such constructive, detailed and informative review which has assisted us in improving our manuscript.
Citation: https://doi.org/10.5194/egusphere-2023-11-AC1
-
RC2: 'Comment on egusphere-2023-11', Anonymous Referee #2, 03 Jun 2023
This paper presents an approach aimed at generating Land Surface Temperature (LST) data at a 1 m resolution using Landsat LST images. The authors establish a correlation between the 900-m resolution LST (derived from the original Landsat LST band) and the 30-m resolution Landsat visible and NIR bands. They also conduct a comparative analysis of the RLS, RF, and SVM algorithms, and compare the performance of regression and machine learning methods (RLS, RF, and SVM) with traditional approaches (DisTrad and TsHARP).
However, I have significant reservations regarding this paper and cannot recommend it for publication based on several key aspects:
Firstly, the LST output does not meet the expected quality standards. The results presented in Figure 5 show significant pixelation and do not exhibit the desired level of detail expected from a 1-meter resolution.
Secondly, the rationale behind establishing a relationship between the visible and NIR bands with LST for the downscaling process seems questionable. The justification for this approach requires further clarification and justification.
Thirdly, the experiment area chosen for validation appears to be too limited in scope. This raises concerns about the generalizability of the findings and the robustness of the proposed approach.
Lastly, the figures presented in the manuscript, such as Figure 2, Figure 3, and Figure 4, fall short of the expected publication quality.
Considering these concerns, I cannot support the publication of this paper in its current form.
Citation: https://doi.org/10.5194/egusphere-2023-11-RC2 -
AC2: 'Reply on RC2', Mario Mhawej, 22 Jun 2023
This paper presents an approach aimed at generating Land Surface Temperature (LST) data at a 1 m resolution using Landsat LST images. The authors establish a correlation between the 900-m resolution LST (derived from the original Landsat LST band) and the 30-m resolution Landsat visible and NIR bands. They also conduct a comparative analysis of the RLS, RF, and SVM algorithms, and compare the performance of regression and machine learning methods (RLS, RF, and SVM) with traditional approaches (DisTrad and TsHARP).
However, I have significant reservations regarding this paper and cannot recommend it for publication based on several key aspects:
Firstly, the LST output does not meet the expected quality standards. The results presented in Figure 5 show significant pixelation and do not exhibit the desired level of detail expected from a 1-meter resolution.
We would like to thank you for this comment. To describe more our approach, below is an example:
When three roofs are seen from space, the RGBN values of each specific roof are retained. These values do not directly provide the LST values as it would be not accurate nor physically-based and this is due to diverse factors we all aware of. What we did instead is we took the main LST value at a higher spatial resolution, namely 30-m, and "adjust" it on a subpixel level based on the 1-m RGBN values using multiple statistical and ML based approaches. As a result, in our approach the coarser resolution LST values were used and slightly enhanced to capture the change in roof LST values between large pixels.
If a land-cover or an emissivity-based based approach is used instead, the three roofs would fall under same class and thus having the same emissivity and class value. Below a descriptive table.
Roof #1
Roof #2
Roof #3
Number of classes
Our RGB approach
30 R, 40 G, 70 B, 110 NIR
130 R, 50 G, 80 B,
21 NIR
30 R, 40 G, 69 B,
111 NIR
3
Land cover or emissivity-based approach
One value for the roof: for instance, 0.2
Same value for the roof as it is in the same land cover type: for instance, 0.2
Same value for the roof as it is in the same land cover type: for instance, 0.2
1
As a result, the newly generated 1-m HSR-GEE is based mainly on 30-m L8 LST and "tuned" with the RGBN values. Because of that the sudden change and discontinuity in spatial patterns in the 30-m L8 is transferred into the 1-m HSR product (a smoother pixels’ transition can be seen for instance in Figure 2 due to a smoother transition in the L8 LST product’s pixels). This "pixelation" is mostly due to cloud coverage and can be overcome by applying a resampling/focal/neighborhood analysis approach at the end of the process but it would erode the data and thus was not recommended in many downscaling studies. Another option would be to include a data imputation technique which is beyond this research. Still, it can be tested in our future research to enable a smoother transition between 1-m LST pixels.
It is important to note however that this did not interfere with the product quality as the generated outputs were validated following the best practices for satellite-derived LST product validation described in Guillevic et al. (2018) using airborne TIR images and yielding very promising accuracy.
Secondly, the rationale behind establishing a relationship between the visible and NIR bands with LST for the downscaling process seems questionable. The justification for this approach requires further clarification and justification.
We would like to thank you for this interesting comment. As discussed in the Introduction section, two main approaches are used for the downscaling purposes. The first is physical-based and the second is statistical-based. Both, particularly the statistical-based approaches are widely available in the literature. These later have used different kind of variables, even the most remote or indirectly related to LST (such as population, income level, etc.). Still, it is important to note here that the produced HSR LST is based on both L8 LST and HSR RGBN images, not only HSR RGBN.
If the reviewer is suggesting to further assess bands and produce an index such the Normalized Difference Vegetation Index (NDVI) or any other index which could somehow depict a physical properties of land features, then, it is important to note the main objective of this paper. The aim of this paper is to use the freely-available 1-m NAIP multi-spectral images in the downscaling of openly available low resolution LST information. Such downscaling process would benefit multiple disciplines in science, including urban-, environmental- and agricultural-based studies, to name a few. Even when using commercial HSR images which come with limited number of bands, and where in some cases the addition of bands would generate more fees, in this study, we tried to retrieve an approach that uses only limited number of bands, namely red, green, blue and NIR, which should be available in all HSR images in different sensors, coupled with the freely available L8 LST images. Of course, including the short-wave infrared (SWIR) band, particularly for urban-based studies is always recommended, but it would generate further limitations for users to implement the HSR-GEE algorithm if the SWIR band is not available. The same goes for generating albedo, an interesting variable for LST calculation, but requiring 5 different bands in Landsat-8, very difficult to retrieve in HSR images.
To summarize, authors of this paper have first intended to include some of the most common indices or parameters, such as NDVI, LAI, or albedo, to cover structural and moisture contents but due to the availability of only four bands (i.e., RGBN bands) and because of the sustainability of the system, the HSR-GEE algorithm was proposed using a widely acceptable statistical-based approach and generating HSR LST based only on HSR RGBN bands and L8 LST values. Further studies could try to include other bands and factors whenever possible as well as including in the analysis and interpretation the potential effect of thermal anisotropy, and more importantly, differences in albedo for different surfaces. Again, thank you for pointing this matter out.
Thirdly, the experiment area chosen for validation appears to be too limited in scope. This raises concerns about the generalizability of the findings and the robustness of the proposed approach.
We would like to thank you for this comment. Concerning the downscaling performance, we did investigate our products following the best practices for satellite-derived LST product validation described in Guillevic et al. (2018), with both 1) validations and 2) inter-evaluation approaches used.
Validations of the HSR-GEE outputs were carried out using airborne TIR images in three different locations and under different conditions. The used TIR images were based on the availability of such images, which are usually scare and very limited. Anyhow, accuracies were very promising, with SVM having the best performance with a RMSE of 1.68°C and a MAE of 1.33°C, followed by RLS with a RMSE of 2.24°C and a MAE of 1.92°C. RF appears to be underperforming in comparison to the two other approaches, with a MAE of 2.53°C.
Furthermore, inter-evaluations were made in comparison to other widely used and validated approaches such as DisTrad and TsHARP and over the three study regions, where HSR GEE main products outperformed them. As a result, the proposed HSR GEE system is transferable over the CONUS and robust.
As HSR-GEE being open-source, it can be always be improved to accommodate any addition that might increase its portability and generalizability. Still, with such promising accuracy retrieved and while HSR-GEE is calculating weighting factors for the downscaling in a dynamic matter and over each location/date, this system can be considered accurate, transferable and portable.
Guillevic, P., F. Göttsche, J. Nickeson, G. Hulley, D. Ghent, Y. Yu, I. Trigo et al. "Land surface temperature product validation best practice protocol. Version 1.1." Best Practice for Satellite-Derived Land Product Validation 60 (2018).
Lastly, the figures presented in the manuscript, such as Figure 2, Figure 3, and Figure 4, fall short of the expected publication quality.
We would like to thank you for mentioning this out. We improved these Figures as suggested while aiming to 1) use even color gradients, 2) avoid red-green combinations and 3) better black and white readability as suggested by the journal guide for authors. The improved images can be found in the updated version of the manuscript.
We would like to thank you for taking the required time to provide us with such detailed and insightful review which has greatly assisted us in improving our manuscript.
Citation: https://doi.org/10.5194/egusphere-2023-11-AC2
-
AC2: 'Reply on RC2', Mario Mhawej, 22 Jun 2023
Status: closed
-
CEC1: 'Comment on egusphere-2023-11', Juan Antonio Añel, 06 Apr 2023
Dear authors,
Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".
https://www.geoscientific-model-development.net/policies/code_and_data_policy.htmlThis is especially unfortunate, as the Topical Editor of your manuscript has already requested twice you move the code to one of the suitable repositories, as we can not accept either the storage in Google Cloud or that it is necessary to log in to get access to it.
Currently, the situation with your manuscript is irregular, as it should not have been accepted for Discussions as it does not comply with our policy. Therefore we will have to reject your manuscript for publication unless you solve this situation in a prompt manner.
In this way, please, publish your code in one of the appropriate repositories, and reply to this comment with the relevant information (link and DOI) as soon as possible. Also, if the topical editor requests you a reviewed version of your manuscript, you must include in it the modified 'Code and Data Availability' section, including the DOI of the code.
I must emphasize again that failing to comply with this request will result in the rejection of your manuscript for publication.
Juan A. Añel
Geosci. Model Dev.
Citation: https://doi.org/10.5194/egusphere-2023-11-CEC1 -
CC1: 'Reply on CEC1', Mario Mhawej, 13 Apr 2023
Dear Dr. Añel,
Many thanks for your comment!
Kindly note that indeed the Topical Editor has previously requested to change the code and we were able to reply to him and explain the situation. As we completely understand the "code and data policy" of the journal, it is important to note that the issue lies in the Google Earth Engine (GEE) and it is not from our end. More precisely, HSR-GEE was developed based on the Java Script language which is fully built on the GEE platform to save on time and resources for users, as it uses the massive computation power of GEE along its very large databases. The only drawback is that GEE requests to have a google-based username to run the code (the source can be accessed freely but cannot run without credentials). Still, even as authors and developers of the HSR-GEE we do not have any access whatsoever to who has used the provided link or which credentials were used to access the code. As a result, reviewers anonymity is reserved.
Anyhow, we have tried to build an application from the HSR-GEE to overcome the fact that users need an account the run the code, which has required a lot of resources and time, only to adhere to the requested policy as suggested. Now, anyone can run the HSR-GEE application from: https://mariomhawej.users.earthengine.app/view/hsr-gee
It is important to note, however, that this application has its own limitations, related to the user limitation capacity at the GEE, which has added few constraints on the full usage of the HSR-GEE system. This include:
- Limited study area at each run;
- Longer processing time;
- Capacity limit issue.
We will be contacting the GEE team to inquire more about how to overcome these issues. In the meantime, we believe reviewers and any interested party can access and run freely the HSR-GEE application (https://mariomhawej.users.earthengine.app/view/hsr-gee) or via credentials to access and run the code wihtout the application as mentioned in the manuscript. Both source codes can be accessed on https://github.com/mariomhawej/HSR-GEE/
We hope this helps,
Kind regards,
Mario M.
Citation: https://doi.org/10.5194/egusphere-2023-11-CC1 -
CEC2: 'Reply on CC1', Juan Antonio Añel, 13 Apr 2023
Dear authors,
I understand your explanation; however, it was your choice to use the GEE system for your work, and despite choosing a system that is not open, you are trying to publish your work in a journal with a strict policy on access to the code.
Also, I think that you have not understood our request adequately. It is not that we need to be able to run the code but that we need it stored in an open repository. You point to one in GitHub; however, again, our policy is very clear about the fact that GitHub is not acceptable for scientific publishing. I would ask you to read the code and data policies of the journal and comply with them, as you want that your work is considered for publication in Geosc. Model Dev.
GitHub itself instructs authors to use other alternatives for long-term archival and publishing, such as Zenodo (GitHub provides a direct way to copy your project to a Zenodo repository). Therefore, please, publish your code in one of the appropriate repositories.
Also, in the GitHub repository, there is no license listed. If you do not include a license, the code continues to be your property and can not be used by others. Therefore, when uploading the code to the repository, you could want to choose a free software/open-source (FLOSS) license. We recommend the GPLv3. You only need to include the file 'https://www.gnu.org/licenses/gpl-3.0.txt' as LICENSE.txt with your code. Also, you can choose other options: GPLv2, Apache License, MIT License, etc.
Please, reply as soon as possible to this comment with the link to the new repository (and DOI) so that it is available for the peer-review process, as it should be.
I insist that if you continue failing to comply with the requirements of the journal, your manuscript will be rejected.
Juan A. Añel
Geosci. Model Dev. Executive Editor
Citation: https://doi.org/10.5194/egusphere-2023-11-CEC2 -
CC2: 'Reply on CEC2', Mario Mhawej, 14 Apr 2023
Dear Dr. Añel,
Many thanks for understanding and for your continued support!
Here is the script DOI: https://doi.org/10.5281/zenodo.7828718
Many thanks,
Kind regards,
Mario M.Citation: https://doi.org/10.5194/egusphere-2023-11-CC2
-
CC2: 'Reply on CEC2', Mario Mhawej, 14 Apr 2023
-
CC1: 'Reply on CEC1', Mario Mhawej, 13 Apr 2023
-
RC1: 'Comment on egusphere-2023-11', Anonymous Referee #1, 28 May 2023
This paper aimed to develop a method to generate 1 m resolution Land Surface Temperature (LST) data from Landsat LST images. They build a relationship between 900-m resolution LST (upscaling for original Landsat LST band) with 30-m resolution Landsat visible bands and NIR band. Then, they downscaled the Landsat original LST band to 1 m by transferring this relationship via RLS, RF, and SVM methods and the 1 m NAIP visible and NIR bands. They also compared regression and machine learning (RLS, RF, and SVM) methods with traditional methods (DisTrad and TsHARP). Although, the regression method, especially the RLS showed the best accuracy when compared with airborne LST data in three validation areas. I still have some concerns before this manuscript is recommended for publication.
Major comments:
- The LST bands in GEE are resampled to 30 m, in fact, the original resolution of Thermal infrared 1 bands for Landsat 8 TIR is 100 m, when you used 900 m LST as an input variable for the regression and machine learning, the scale is 9 not 30.
- The principle between object reflection and emission is different. In the visible bands, objects reflect different amounts of light depending on their color and composition. In the NIR bands, they absorb or reflect light based on their structural and moisture content. While the emission ability is dependent on the thermal properties of objects. Meanwhile, The spatial heterogeneity of the temperature of ground objects is not the same as its reflectance band. Why can visible light and near-infrared bands be used for downscaling? The authors should make it clear.
- The spatial pattern of temperature in Figure 1 and Figure 5 (in Clayville) do not show high quality, the temperature change in adjacent areas is not continuous. For example, in the middle of the right part of Figure 1, the blue is an obvious sudden change in temperature, and in Figure 5, the temperature extracted with HSR RSL method in claryville (bottom figure) shows an obvious cross line. The sudden change and discontinuity in spatial patterns should be improved.
Citation: https://doi.org/10.5194/egusphere-2023-11-RC1 -
AC1: 'Reply on RC1', Mario Mhawej, 22 Jun 2023
This paper aimed to develop a method to generate 1 m resolution Land Surface Temperature (LST) data from Landsat LST images. They build a relationship between 900-m resolution LST (upscaling for original Landsat LST band) with 30-m resolution Landsat visible bands and NIR band. Then, they downscaled the Landsat original LST band to 1 m by transferring this relationship via RLS, RF, and SVM methods and the 1 m NAIP visible and NIR bands. They also compared regression and machine learning (RLS, RF, and SVM) methods with traditional methods (DisTrad and TsHARP). Although, the regression method, especially the RLS showed the best accuracy when compared with airborne LST data in three validation areas. I still have some concerns before this manuscript is recommended for publication.
Major comments:
- The LST bands in GEE are resampled to 30 m, in fact, the original resolution of Thermal infrared 1 bands for Landsat 8 TIR is 100 m, when you used 900 m LST as an input variable for the regression and machine learning, the scale is 9 not 30.
We would like to thank you for this comment. Indeed, the used USGS Landsat 8 Level 2, Collection 2, Tier 1 TIR product is resampled to 30 meters to match multispectral bands. The scale is 9 instead of 30. The manuscript was improved accordingly.
- The principle between object reflection and emission is different. In the visible bands, objects reflect different amounts of light depending on their color and composition. In the NIR bands, they absorb or reflect light based on their structural and moisture content. While the emission ability is dependent on the thermal properties of objects. Meanwhile, The spatial heterogeneity of the temperature of ground objects is not the same as its reflectance band. Why can visible light and near-infrared bands be used for downscaling? The authors should make it clear.
We would like to thank you for this interesting comment. As discussed in the Introduction section, two main approaches are used for the downscaling purposes. The first is physical-based and the second is statistical-based. Both, particularly the statistical-based approaches are widely available in the literature. These later have used different kind of variables, even the most remote or indirectly related to LST (such as population, income level, etc.). Still, it is important to note here that the produced HSR LST is based on both L8 LST and HSR RGBN images, not only HSR RGBN.
If the reviewer is suggesting to further assess bands and produce an index such the Normalized Difference Vegetation Index (NDVI) or any other index which could somehow depict a physical properties of land features, then, it is important to note the main objective of this paper. The aim of this paper is to use the freely-available 1-m NAIP multi-spectral images in the downscaling of openly available low resolution LST information. Such downscaling process would benefit multiple disciplines in science, including urban-, environmental- and agricultural-based studies, to name a few. Even when using commercial HSR images which come with limited number of bands, and where in some cases the addition of bands would generate more fees, in this study, we tried to retrieve an approach that uses only limited number of bands, namely red, green, blue and NIR, which should be available in all HSR images in different sensors, coupled with the freely available L8 LST images. Of course, including the short-wave infrared (SWIR) band, particularly for urban-based studies is always recommended, but it would generate further limitations for users to implement the HSR-GEE algorithm if the SWIR band is not available. The same goes for generating albedo, an interesting variable for LST calculation, but requiring 5 different bands in Landsat-8, very difficult to retrieve in HSR images.
To summarize, authors of this paper have first intended to include some of the most common indices or parameters, such as NDVI, LAI, or albedo, to cover structural and moisture contents but due to the availability of only four bands (i.e., RGBN bands) and because of the sustainability of the system, the HSR-GEE algorithm was proposed using a widely acceptable statistical-based approach and generating HSR LST based only on HSR RGBN bands and L8 LST values. Further studies could try to include other bands and factors whenever possible as well as including in the analysis and interpretation the potential effect of thermal anisotropy, and more importantly, differences in albedo for different surfaces. Again, thank you for pointing this matter out.
- The spatial pattern of temperature in Figure 1 and Figure 5 (in Clayville) do not show high quality, the temperature change in adjacent areas is not continuous. For example, in the middle of the right part of Figure 1, the blue is an obvious sudden change in temperature, and in Figure 5, the temperature extracted with HSR RSL method in claryville (bottom figure) shows an obvious cross line. The sudden change and discontinuity in spatial patterns should be improved.
We would like to thank you for this comment. To describe more our approach, below is an example:
When three roofs are seen from space, the RGBN values of each specific roof are retained. These values do not directly provide the LST values as it would be not accurate nor physically-based and this is due to diverse factors we all aware of. What we did instead is we took the main LST value at a higher spatial resolution, namely 30-m, and "adjust" it on a subpixel level based on the 1-m RGBN values using multiple statistical and ML based approaches. As a result, in our approach the coarser resolution LST values were used and slightly enhanced to capture the change in roof LST values between large pixels.
If a land-cover or an emissivity-based based approach is used instead, the three roofs would fall under same class and thus having the same emissivity and class value. Below a descriptive table.
Roof #1
Roof #2
Roof #3
Number of classes
Our RGB approach
30 R, 40 G, 70 B, 110 NIR
130 R, 50 G, 80 B,
21 NIR
30 R, 40 G, 69 B,
111 NIR
3
Land cover or emissivity-based approach
One value for the roof: for instance, 0.2
Same value for the roof as it is in the same land cover type: for instance, 0.2
Same value for the roof as it is in the same land cover type: for instance, 0.2
1
As a result, the newly generated 1-m HSR-GEE is based mainly on 30-m L8 LST and "tuned" with the RGBN values. Because of that the sudden change and discontinuity in spatial patterns in the 30-m L8 is transferred into the 1-m HSR product (a smoother pixels’ transition can be seen for instance in Figure 2 due to a smoother transition in the L8 LST product’s pixels). This "pixelation" is mostly due to cloud coverage and can be overcome by applying a resampling/focal/neighborhood analysis approach at the end of the process but it would erode the data and thus was not recommended in many downscaling studies. Another option would be to include a data imputation technique which is beyond this research. Still, it can be tested in our future research to enable a smoother transition between 1-m LST pixels.
It is important to note however that this did not interfere with the product quality as the generated outputs were validated following the best practices for satellite-derived LST product validation described in Guillevic et al. (2018) using airborne TIR images and yielding very promising accuracy.
We would like to thank you for taking the required time to provide us with such constructive, detailed and informative review which has assisted us in improving our manuscript.
Citation: https://doi.org/10.5194/egusphere-2023-11-AC1
-
RC2: 'Comment on egusphere-2023-11', Anonymous Referee #2, 03 Jun 2023
This paper presents an approach aimed at generating Land Surface Temperature (LST) data at a 1 m resolution using Landsat LST images. The authors establish a correlation between the 900-m resolution LST (derived from the original Landsat LST band) and the 30-m resolution Landsat visible and NIR bands. They also conduct a comparative analysis of the RLS, RF, and SVM algorithms, and compare the performance of regression and machine learning methods (RLS, RF, and SVM) with traditional approaches (DisTrad and TsHARP).
However, I have significant reservations regarding this paper and cannot recommend it for publication based on several key aspects:
Firstly, the LST output does not meet the expected quality standards. The results presented in Figure 5 show significant pixelation and do not exhibit the desired level of detail expected from a 1-meter resolution.
Secondly, the rationale behind establishing a relationship between the visible and NIR bands with LST for the downscaling process seems questionable. The justification for this approach requires further clarification and justification.
Thirdly, the experiment area chosen for validation appears to be too limited in scope. This raises concerns about the generalizability of the findings and the robustness of the proposed approach.
Lastly, the figures presented in the manuscript, such as Figure 2, Figure 3, and Figure 4, fall short of the expected publication quality.
Considering these concerns, I cannot support the publication of this paper in its current form.
Citation: https://doi.org/10.5194/egusphere-2023-11-RC2 -
AC2: 'Reply on RC2', Mario Mhawej, 22 Jun 2023
This paper presents an approach aimed at generating Land Surface Temperature (LST) data at a 1 m resolution using Landsat LST images. The authors establish a correlation between the 900-m resolution LST (derived from the original Landsat LST band) and the 30-m resolution Landsat visible and NIR bands. They also conduct a comparative analysis of the RLS, RF, and SVM algorithms, and compare the performance of regression and machine learning methods (RLS, RF, and SVM) with traditional approaches (DisTrad and TsHARP).
However, I have significant reservations regarding this paper and cannot recommend it for publication based on several key aspects:
Firstly, the LST output does not meet the expected quality standards. The results presented in Figure 5 show significant pixelation and do not exhibit the desired level of detail expected from a 1-meter resolution.
We would like to thank you for this comment. To describe more our approach, below is an example:
When three roofs are seen from space, the RGBN values of each specific roof are retained. These values do not directly provide the LST values as it would be not accurate nor physically-based and this is due to diverse factors we all aware of. What we did instead is we took the main LST value at a higher spatial resolution, namely 30-m, and "adjust" it on a subpixel level based on the 1-m RGBN values using multiple statistical and ML based approaches. As a result, in our approach the coarser resolution LST values were used and slightly enhanced to capture the change in roof LST values between large pixels.
If a land-cover or an emissivity-based based approach is used instead, the three roofs would fall under same class and thus having the same emissivity and class value. Below a descriptive table.
Roof #1
Roof #2
Roof #3
Number of classes
Our RGB approach
30 R, 40 G, 70 B, 110 NIR
130 R, 50 G, 80 B,
21 NIR
30 R, 40 G, 69 B,
111 NIR
3
Land cover or emissivity-based approach
One value for the roof: for instance, 0.2
Same value for the roof as it is in the same land cover type: for instance, 0.2
Same value for the roof as it is in the same land cover type: for instance, 0.2
1
As a result, the newly generated 1-m HSR-GEE is based mainly on 30-m L8 LST and "tuned" with the RGBN values. Because of that the sudden change and discontinuity in spatial patterns in the 30-m L8 is transferred into the 1-m HSR product (a smoother pixels’ transition can be seen for instance in Figure 2 due to a smoother transition in the L8 LST product’s pixels). This "pixelation" is mostly due to cloud coverage and can be overcome by applying a resampling/focal/neighborhood analysis approach at the end of the process but it would erode the data and thus was not recommended in many downscaling studies. Another option would be to include a data imputation technique which is beyond this research. Still, it can be tested in our future research to enable a smoother transition between 1-m LST pixels.
It is important to note however that this did not interfere with the product quality as the generated outputs were validated following the best practices for satellite-derived LST product validation described in Guillevic et al. (2018) using airborne TIR images and yielding very promising accuracy.
Secondly, the rationale behind establishing a relationship between the visible and NIR bands with LST for the downscaling process seems questionable. The justification for this approach requires further clarification and justification.
We would like to thank you for this interesting comment. As discussed in the Introduction section, two main approaches are used for the downscaling purposes. The first is physical-based and the second is statistical-based. Both, particularly the statistical-based approaches are widely available in the literature. These later have used different kind of variables, even the most remote or indirectly related to LST (such as population, income level, etc.). Still, it is important to note here that the produced HSR LST is based on both L8 LST and HSR RGBN images, not only HSR RGBN.
If the reviewer is suggesting to further assess bands and produce an index such the Normalized Difference Vegetation Index (NDVI) or any other index which could somehow depict a physical properties of land features, then, it is important to note the main objective of this paper. The aim of this paper is to use the freely-available 1-m NAIP multi-spectral images in the downscaling of openly available low resolution LST information. Such downscaling process would benefit multiple disciplines in science, including urban-, environmental- and agricultural-based studies, to name a few. Even when using commercial HSR images which come with limited number of bands, and where in some cases the addition of bands would generate more fees, in this study, we tried to retrieve an approach that uses only limited number of bands, namely red, green, blue and NIR, which should be available in all HSR images in different sensors, coupled with the freely available L8 LST images. Of course, including the short-wave infrared (SWIR) band, particularly for urban-based studies is always recommended, but it would generate further limitations for users to implement the HSR-GEE algorithm if the SWIR band is not available. The same goes for generating albedo, an interesting variable for LST calculation, but requiring 5 different bands in Landsat-8, very difficult to retrieve in HSR images.
To summarize, authors of this paper have first intended to include some of the most common indices or parameters, such as NDVI, LAI, or albedo, to cover structural and moisture contents but due to the availability of only four bands (i.e., RGBN bands) and because of the sustainability of the system, the HSR-GEE algorithm was proposed using a widely acceptable statistical-based approach and generating HSR LST based only on HSR RGBN bands and L8 LST values. Further studies could try to include other bands and factors whenever possible as well as including in the analysis and interpretation the potential effect of thermal anisotropy, and more importantly, differences in albedo for different surfaces. Again, thank you for pointing this matter out.
Thirdly, the experiment area chosen for validation appears to be too limited in scope. This raises concerns about the generalizability of the findings and the robustness of the proposed approach.
We would like to thank you for this comment. Concerning the downscaling performance, we did investigate our products following the best practices for satellite-derived LST product validation described in Guillevic et al. (2018), with both 1) validations and 2) inter-evaluation approaches used.
Validations of the HSR-GEE outputs were carried out using airborne TIR images in three different locations and under different conditions. The used TIR images were based on the availability of such images, which are usually scare and very limited. Anyhow, accuracies were very promising, with SVM having the best performance with a RMSE of 1.68°C and a MAE of 1.33°C, followed by RLS with a RMSE of 2.24°C and a MAE of 1.92°C. RF appears to be underperforming in comparison to the two other approaches, with a MAE of 2.53°C.
Furthermore, inter-evaluations were made in comparison to other widely used and validated approaches such as DisTrad and TsHARP and over the three study regions, where HSR GEE main products outperformed them. As a result, the proposed HSR GEE system is transferable over the CONUS and robust.
As HSR-GEE being open-source, it can be always be improved to accommodate any addition that might increase its portability and generalizability. Still, with such promising accuracy retrieved and while HSR-GEE is calculating weighting factors for the downscaling in a dynamic matter and over each location/date, this system can be considered accurate, transferable and portable.
Guillevic, P., F. Göttsche, J. Nickeson, G. Hulley, D. Ghent, Y. Yu, I. Trigo et al. "Land surface temperature product validation best practice protocol. Version 1.1." Best Practice for Satellite-Derived Land Product Validation 60 (2018).
Lastly, the figures presented in the manuscript, such as Figure 2, Figure 3, and Figure 4, fall short of the expected publication quality.
We would like to thank you for mentioning this out. We improved these Figures as suggested while aiming to 1) use even color gradients, 2) avoid red-green combinations and 3) better black and white readability as suggested by the journal guide for authors. The improved images can be found in the updated version of the manuscript.
We would like to thank you for taking the required time to provide us with such detailed and insightful review which has greatly assisted us in improving our manuscript.
Citation: https://doi.org/10.5194/egusphere-2023-11-AC2
-
AC2: 'Reply on RC2', Mario Mhawej, 22 Jun 2023
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 632 | 374 | 35 | 1,041 | 15 | 16 |
- HTML: 632
- PDF: 374
- XML: 35
- Total: 1,041
- BibTeX: 15
- EndNote: 16
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1