the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
High-resolution and Multitemporal Impervious Surface Mapping in the Lancang-Mekong Basin with Google Earth Engine
Abstract. High-resolution and multitemporal impervious surface maps on large scales are crucial for environmental and socioeconomic studies. However, recently available multitemporal impervious surface maps of the Lancang-Mekong basin were limited at 30-m resolution with considerably low accuracy. Hence, the development of up-to-date, accurate, and multitemporal impervious surface maps with the 10-m resolution is urgently needed. In this article, a machine learning framework is demonstrated by fusing Sentinel-1 Synthetic Aperture Radar images and Sentinel-2 Multispectral Sensor images to map and study the annual dynamics of impervious surfaces in the Lancang-Mekong basin from 2016 to 2021 facilitated by Google Earth Engine. Moreover, a train sample migration strategy is proposed to automate impervious surface mapping for various time periods eliminating the need to collect additional train samples from this vast study area. Finally, qualitative and quantitative assessments are conducted using test samples from Google Earth and four existing state-of-the-art datasets. The result shows that the overall accuracy and Kappa of the final impervious surface maps range from 91.45 % to 92.44 % and 0.829 to 0.848, respectively, which demonstrates the feasibility and reliability of the proposed method and results. The LMISD is freely available from https://doi.org/10.5281/zenodo.6968739 (Sun et al., 2022).
- Preprint
(3843 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed
-
RC1: 'Comment on essd-2022-251', Anonymous Referee #1, 04 Sep 2022
The manuscript "High-resolution and Multitemporal Impervious Surface Mapping in the Lancang-Mekong Basin with Google Earth Engine" describes a new IS map for Mekong basin. To this aim, the authors applied visual-interpreted samples to multi-source features derived from Sentinel-1/2 images. The whole framework is routine and some key parameters should be further clarified or discussed. Current accuracy assessment showed better performance than some datasets. However, comparison with the state-of-art datasets is missing, the superiority to the existing works remains unclear. Moreover, the test samples seem to be located mostly in urban and sub-urban areas, leading to incomplete accuracy assessment over rural regions. The structural of manuscript is clear but the writing should be carefully revised. As such, I would suggest a rejection.
Line 70. Many 10-m global datasets have been released, such as the GHS-BUILT-S (Corbane et al., 2021) from EC JRC and the GISA-10m (Huang et al., 2022) from Wuhan University.
Corbane, C., Syrris, V., Sabo, F., Politis, P., Melchiorri, M., Pesaresi, M., Soille, P., and Kemper, T.: Convolutional neural networks for global human settlements mapping from Sentinel-2 satellite imagery, Neural Comput. Appl., 33, 6697–6720, https://doi.org/10.1007/s00521-020-05449-7, 2021.
Huang, X., Yang, J., Wang, W., and Liu, Z.: Mapping 10 m global impervious surface area (GISA-10m) using multi-source geospatial data, Earth Syst. Sci. Data, 14, 3649–3672, https://doi.org/10.5194/essd-14-3649-2022, 2022.
Line 100. I would delete the "support" here.
Line 125. Did you just stack the Sentine-1 data from both "ascending" and "descending" orbit together? This will lead to distortions over mountain areas if you do so.
Line 195. I know these metrices make sense, but you may explain why they were chosen.
Line 199. Could you explain why temporal metrices were only derived from Sentinel-1? There are much more Sentinel-2 images (Figure 2).
Table 2."Number" ->"Dimension".
Line 205. Could you show us the distribution of training samples?
Line 233. How this threshold (SAD > 0.125) was determined?
Line 235. Why a pixel can be regarded as water if its MNDWI is greater than 0.12? It's not a robust method.
Line 240. Without seeing your training samples, I presume many pixels identified as changed may not change from NIS to IS if they are far from IS. Could you do a quick ablation experiment to better demonstrate the effectiveness of the sample migration?
Line 265. The methods (and figures) you described here are similar to that in Li et al., (2015). You should at least cite it.
Li, X., Gong, P., and Liang, L.: A 30-year (1984–2013) record of annual urban dynamics of Beijing City derived from Landsat data, Remote Sens. Environ., 166, 78–90, https://doi.org/10.1016/j.rse.2015.06.007, 2015.
Line 272. If your results follow the assumption that transition from IS to NIS is rare, you may reduce the results to a single band where pixel indicates the time when IS was first detected, instead of putting annual data into separate bands.
Figure 6. It seems that most of your IS test samples locate in or near cities. Could you provide the ISA density around the IS samples? Courtesy of higher spatial resolution, buildings and roads in rural regions can be better delineated in Sentinel images. Therefore, accuracy assessment over rural regions for 10-m IS mapping is important.
Line 302. Are there new (or interesting) findings from your results?
Line 328. "0:05" ->"0.05"?
Table 5. It's would be more interesting to compare your results with GHS and GISA-10m (I mentioned above). They are both 10-m thematic mappings, same as you did here. I strongly suggest you to do so.
Line 377. In fact, according to your results (Table 5), the ESA-2020 achieved higher PA in all cities, indicating that ESA-2020 has less omissions. This is contradictory with the statement here that ESA-2020 ignores buildings in rural areas. Could you explain it?
Citation: https://doi.org/10.5194/essd-2022-251-RC1 -
CC1: 'Reply on RC1', Aizhu Zhang, 17 Sep 2022
The manuscript "High-resolution and Multitemporal Impervious Surface Mapping in the Lancang-Mekong Basin with Google Earth Engine" describes a new IS map for Mekong basin. To this aim, the authors applied visual-interpreted samples to multi-source features derived from Sentinel-1/2 images. The whole framework is routine and some key parameters should be further clarified or discussed. Current accuracy assessment showed better performance than some datasets. However, comparison with the state-of-art datasets is missing, the superiority to the existing works remains unclear. Moreover, the test samples seem to be located mostly in urban and sub-urban areas, leading to incomplete accuracy assessment over rural regions. The structural of manuscript is clear but the writing should be carefully revised. As such, I would suggest a rejection.
Line 70. Many 10-m global datasets have been released, such as the GHS-BUILT-S (Corbane et al., 2021) from EC JRC and the GISA-10m (Huang et al., 2022) from Wuhan University.
Corbane, C., Syrris, V., Sabo, F., Politis, P., Melchiorri, M., Pesaresi, M., Soille, P., and Kemper, T.: Convolutional neural networks for global human settlements mapping from Sentinel-2 satellite imagery, Neural Comput. Appl., 33, 6697–6720, https://doi.org/10.1007/s00521-020-05449-7, 2021.
Huang, X., Yang, J., Wang, W., and Liu, Z.: Mapping 10 m global impervious surface area (GISA-10m) using multi-source geospatial data, Earth Syst. Sci. Data, 14, 3649–3672, https://doi.org/10.5194/essd-14-3649-2022, 2022.
Response: Thanks for your suggestions. We noted the release of the global 10m impervious surface products, but (1) the globally shared samples and methods may not yield good results for a special region like the Lancang-Mekong Basin, and (2) these products are only available in one phase and do not meet the requirements of dynamic analysis. For this reason, we conducted the development of a 10m multi-temporal impervious surface for the Lancang-Mekong Basin.
Line 100. I would delete the "support" here.
Response: Thanks, we will take this comment.
Line 125. Did you just stack the Sentine-1 data from both "ascending" and "descending" orbit together? This will lead to distortions over mountain areas if you do so.
Response: Our experimental results show that after combining the mask of DEM data, the direct stacking of the two does not have much effect on the misclassification of the mountains.
Line 195. I know these metrices make sense, but you may explain why they were chosen.
Response: The 7*7 size was obtained based on existing studies and experiments.
Line 199. Could you explain why temporal metrices were only derived from Sentinel-1? There are much more Sentinel-2 images (Figure 2).
Response: The original Sentinel-2 images are more than the Sentinel-1 images, but there are many clouds and rain in the Lancang-Mekong Basin, resulting in poor quality of some optical images. After we eliminate this part, the number of Sentinel-2 images is not enough to calculate the temporal features.
Table 2."Number" ->"Dimension".
Response: Thanks, we will take this comment.
Line 205. Could you show us the distribution of training samples?
Response: The samples were obtained by random sampling and manually divided into sample polygons of the same type around them.
Line 233. How this threshold (SAD > 0.125) was determined?
Response: We experimentally obtained the division thresholds by selecting a number of sample points, including both changing and non-changing classes.
Line 235. Why a pixel can be regarded as water if its MNDWI is greater than 0.12? It's not a robust method.
Response: Thanks, this is indeed not a robust approach, but we still use it in order to reduce the computational effort.
Line 240. Without seeing your training samples, I presume many pixels identified as changed may not change from NIS to IS if they are far from IS. Could you do a quick ablation experiment to better demonstrate the effectiveness of the sample migration?
Response: The ablation experiment at large-scale is rather complicated, and we can only give a comparison of the precision in some areas. Without sample migration, the overall precision of the results in 2021 will decrease from 91.45% to 89.86%.
Line 265. The methods (and figures) you described here are similar to that in Li et al., (2015). You should at least cite it.
Li, X., Gong, P., and Liang, L.: A 30-year (1984–2013) record of annual urban dynamics of Beijing City derived from Landsat data, Remote Sens. Environ., 166, 78–90, https://doi.org/10.1016/j.rse.2015.06.007, 2015.
Response: Thanks, we should cite it here.
Line 272. If your results follow the assumption that transition from IS to NIS is rare, you may reduce the results to a single band where pixel indicates the time when IS was first detected, instead of putting annual data into separate bands.
Response: Thanks, this is a good way to reduce the dataset.
Figure 6. It seems that most of your IS test samples locate in or near cities. Could you provide the ISA density around the IS samples? Courtesy of higher spatial resolution, buildings and roads in rural regions can be better delineated in Sentinel images. Therefore, accuracy assessment over rural regions for 10-m IS mapping is important.
Response: We obtained the validation samples by stratified sampling, and although the samples are less at rural areas, as a whole, the representativeness is sufficient. We will use a detailed comparison to illustrate the effect in rural areas.
Line 302. Are there new (or interesting) findings from your results?
Response: I am sorry to say that the main purpose of this paper is to propose mapping methods for large scale impervious surfaces in the Lancang-Mekong Basin and to develop a usable product. But no effort was spent on analyzing the new (or interesting) findings.
Line 328. "0:05" ->"0.05"?
Response: Thanks, we have a spelling error here.
Table 5. It's would be more interesting to compare your results with GHS and GISA-10m (I mentioned above). They are both 10-m thematic mappings, same as you did here. I strongly suggest you to do so.
Response: We are trying to compare with these two datasets, but that will take some time. At least for now, visually, we have a better product than GHS, which is too fragmented in the Lancang-Mekong Basin.
Line 377. In fact, according to your results (Table 5), the ESA-2020 achieved higher PA in all cities, indicating that ESA-2020 has less omissions. This is contradictory with the statement here that ESA-2020 ignores buildings in rural areas. Could you explain it?
Response: As you mentioned before, the results of the accuracy validation are strongly related to the distribution of the samples, and the stratified samples are mostly scattered in urban areas, which is the reason why the ESA accuracy is higher than ours. However, in terms of visual effects, ESA has a certain omission in rural areas.
Citation: https://doi.org/10.5194/essd-2022-251-CC1
-
CC1: 'Reply on RC1', Aizhu Zhang, 17 Sep 2022
-
RC2: 'Comment on essd-2022-251', Anonymous Referee #2, 04 Sep 2022
This manuscript provided 10-m resolution impervious surface area data using Sentinel-1 and Sentinel-2 data in Lancang-Mekong Basin. Overall, the novelty of this paper is not qualified for publication in ESSD for the following reasons.
(1) Mapping of impervious surface area at the global scale with 10-m or 30-m resolutions has been widely reported in many studies. The method employed in this paper doesn’t show significant improvement regarding the novelty of mapping approaches as well as the mapped results and influences.
(2) Although the migration strategy of training samples across years highlights in this paper, some concerns may introduce uncertainties in this study. For example, the sampling in 2016 may omit those urbanized regions in 2021, which may bias the classification results (see Table 4). Also, the threshold of SAD is challenging to determine if these pixels have experienced change or not.
(3) The temporal consistency check is not the novelty of this work. This approach was initially proposed in the RSE paper in 2015 (in the case of Beijing). Unfortunately, I didn’t find the citation of this work when they introduced the “temporal consistency check” part, and the overall framework is highly similar to that one.
Title: A 30-year (1984–2013) record of annual urban dynamics of Beijing City derived from Landsat data.
(4) It is unfair to compare the local mapping results with those of global products because it is a space-specific application at the river basin. I think the proposed approach cannot be extended directly to other regions (even the globe).
Citation: https://doi.org/10.5194/essd-2022-251-RC2 -
CC2: 'Reply on RC2', Aizhu Zhang, 17 Sep 2022
This manuscript provided 10-m resolution impervious surface area data using Sentinel-1 and Sentinel-2 data in Lancang-Mekong Basin. Overall, the novelty of this paper is not qualified for publication in ESSD for the following reasons.
(1) Mapping of impervious surface area at the global scale with 10-m or 30-m resolutions has been widely reported in many studies. The method employed in this paper doesn’t show significant improvement regarding the novelty of mapping approaches as well as the mapped results and influences.
Response: In this paper, we propose a framework for fusing multi-source data based on the GEE platform to address the lack of available multi-temporal 10m impervious surface datasets due to multiple factors in the Lancang-Mekong Basin, and finally publish a dataset of impervious surfaces in the Lancang-Mekong Basin for the period 2016-2021.
(2) Although the migration strategy of training samples across years highlights in this paper, some concerns may introduce uncertainties in this study. For example, the sampling in 2016 may omit those urbanized regions in 2021, which may bias the classification results (see Table 4). Also, the threshold of SAD is challenging to determine if these pixels have experienced change or not.
Response: In fact, the rate of renewal of impervious surface categories is much less than the rate of area expansion. That is, there will be the same material and spatial characteristics of impervious surfaces in urbanized areas in 2021 that existed in 2016. Considering this, the ultimate purpose of samples migration is simply to avoid the error that occurs in imaging over many years of images, and if this error does not exist, it is in fact only necessary to use the model trained in 2016 for classification. About the threshold of SAD, we experimentally obtained the division thresholds by selecting a number of sample points, including both changing and non-changing classes.
(3) The temporal consistency check is not the novelty of this work. This approach was initially proposed in the RSE paper in 2015 (in the case of Beijing). Unfortunately, I didn’t find the citation of this work when they introduced the “temporal consistency check” part, and the overall framework is highly similar to that one.
Title: A 30-year (1984–2013) record of annual urban dynamics of Beijing City derived from Landsat data.
Response: Thanks, we should cite it here.
(4) It is unfair to compare the local mapping results with those of global products because it is a space-specific application at the river basin. I think the proposed approach cannot be extended directly to other regions (even the globe).
Response: It is true that this comparison is unfair and that the globally shared samples and methods may not yield good results for a special region like the Lancang-Mekong Basin. However, this is the only method used for comparison, as no other products specific to the LMC Basin exist. It also illustrates the need to develop products for a particular region, which is the point of our proposed methodology.
Citation: https://doi.org/10.5194/essd-2022-251-CC2
-
CC2: 'Reply on RC2', Aizhu Zhang, 17 Sep 2022
-
RC3: 'Comment on essd-2022-251', Christopher Brown, 20 Sep 2022
The authors present a method for extrapolating annualized annotations for binary impervious surface (IS) / non-impervious surface (NIS) into future years when NIS -> IS transitions are assumed to be insignificantly probable. The authors also present a method for converting a short sequence of annual IS / NIS binary classifications into a sequence congruent with the previous assumption regarding transitions. The authors combine these methods in a standard supervised Sentinel-1 / Sentinel-2 random forest classification workflow in Google Earth Engine to produce annual IS / NIS maps of the Lancang-Mekong basin from 2016 to 2021. These maps are compared to a number of existing global LULC products with an IS class, or a class the authors considered comparable, using a dataset created specifically for this purpose. The results of the author's assessment indicate good overall agreement with their sample.
I would begin by asking the authors to carefully copy-review their manuscript: I find that there are grammatical and/or punctuation errors on nearly all of the proof pages. In particular I suggest paying special attention to extraneous usage of "the" before nouns, and general comma usage. I also suggest replacing verbs used in a subjective context such as "excellent" or "state-of-the-art" with the less strongly worded "popular" or "modern." There are also opportunities to unify language e.g. "dynamic monitoring of IS" is more commonly known as "change monitoring," and the authors are not consistent in their wording. I find the structure of the manuscript's narrative generally well paced and direct.
The author's characterization of the Sentinel-1 and Sentinel-2 missions was imprecise. Line 62 asserts Sentinel-1 has a 10m "spatial" resolution, however Sentinel-1's IW spatial resolution is actually 5m x 20m with the GRD product specifically resampled to a pixel resolution of 10m (https://sentinels.copernicus.eu/web/sentinel/missions/sentinel-1/overview/mission-summary). Line 103 discusses Sentinel-2 "image blocks," which is not a Sentinel-2 data product I'm aware of, and lines 103-104 discuss Sentinel-1 and SRTM "scenes," which are not part of the ARD scheme used in delivering either of these products. On line 123, the authors state Sentinel-1's revisit to be 6 days, and while this is true of the instrument, the availability of the data used by the authors (GRD IW) is highly variable between orbits, location, and polarization (https://www.researchgate.net/figure/Sentinel-1-constellation-observation-revisit-and-coverage-scenario-April-2021-10_fig2_351632665).
I have concerns with the expert system introduced by the authors to perform sample migration, that which formed the basis of their strategy to extend their training data. Besides a lack of ablation analysis on the many parameters, Table 4 (which would benefit from NIS column totals) raises doubts about the author's procedure. Considering that a fundamental assumption in the author's framework is that IS -> NIS changes are not permitted, one would expect points migrated to successive years to decrease with time. This is not the case, and e.g. points that were not migrated to 2017 (15,649 total) are in fact migrated to 2018 (16,075), calling into question the validity of this approach since it violates the IS -> NIS assumption between 2017-2018. At the extremes, it would be possible for a point to exist exclusively in 2016 and 2021. I am also concerned with the use of pixel-counting statistics used in Equations 2-5, as it is not stated whether or not the authors account for processing baseline updates in the underlying Sentinel-2 collection. Failing to do so would lead to double-counting and so erroneous results.
I appreciate that the procedural diagrams in Figures 3, 4, and 5 were easy to follow, which is somewhat of a rarity.
I also have concerns with the temporal consistency check introduced by the authors. From first principles, the authors are, in my opinion, suggesting a rather complicated procedure to post-process a length 6 vector with only 64 possible values, 57 of which are actually invalid. The authors do not state the boundary considerations for the sliding window, which is important as it effectively defines the IS change sequence semantics in ambiguous cases like "101010." Furthermore, that such ambiguous sequences are collapsed to "valid" sequences seems like a fallacious assumption in the overall methodology when a "transition" or "low-confidence" classification may be more appropriate.
Little information is given regarding the stratification performed to collect validation data.
I find that Figures 10 and 11 and the corresponding section offer little insight or value into the author's data.
The author's compute area statistics from their maps yet do not offer any confidence intervals or uncertainties. These statistics are all but the bare minimum required to publish estimations of area based on remotely sensed data.
I recommend the authors discuss the potential for change dynamics to effect their agreement statistics with the external data products outside of the sampling year. The authors conclude on Lines 348-349 that similar accuracies to WorldCover-10 2020 in 2021 indicate that the sample migration strategy was successful, but since no baseline was established, it is unclear whether similar performance can be attributable to a combination of class-shift and definition penalizing WorldCover-10, extra training data introduced in re-using points across years, or the nuances of the algorithm introduced by the authors.
One of the biggest potential contributions from this effort would have been the underlying training data sampled from 2016, and I encourage the authors to make this available.
While there is good work in this research effort, I believe the authors should intensely proofread their writing and try to keep their language exact and consistent. I also believe the novelty, the suggested sample migration and temporal smoothing, demands more methodological consideration and emphasis in the manuscript. The algorithms introduced also require more justification to contextualize their effects on the accuracy assessment. The authors should seek existing literature on area estimation to ensure confidences are published with the computed statistics. My overall recommendation is for reconsideration after major revisions.
Citation: https://doi.org/10.5194/essd-2022-251-RC3
Status: closed
-
RC1: 'Comment on essd-2022-251', Anonymous Referee #1, 04 Sep 2022
The manuscript "High-resolution and Multitemporal Impervious Surface Mapping in the Lancang-Mekong Basin with Google Earth Engine" describes a new IS map for Mekong basin. To this aim, the authors applied visual-interpreted samples to multi-source features derived from Sentinel-1/2 images. The whole framework is routine and some key parameters should be further clarified or discussed. Current accuracy assessment showed better performance than some datasets. However, comparison with the state-of-art datasets is missing, the superiority to the existing works remains unclear. Moreover, the test samples seem to be located mostly in urban and sub-urban areas, leading to incomplete accuracy assessment over rural regions. The structural of manuscript is clear but the writing should be carefully revised. As such, I would suggest a rejection.
Line 70. Many 10-m global datasets have been released, such as the GHS-BUILT-S (Corbane et al., 2021) from EC JRC and the GISA-10m (Huang et al., 2022) from Wuhan University.
Corbane, C., Syrris, V., Sabo, F., Politis, P., Melchiorri, M., Pesaresi, M., Soille, P., and Kemper, T.: Convolutional neural networks for global human settlements mapping from Sentinel-2 satellite imagery, Neural Comput. Appl., 33, 6697–6720, https://doi.org/10.1007/s00521-020-05449-7, 2021.
Huang, X., Yang, J., Wang, W., and Liu, Z.: Mapping 10 m global impervious surface area (GISA-10m) using multi-source geospatial data, Earth Syst. Sci. Data, 14, 3649–3672, https://doi.org/10.5194/essd-14-3649-2022, 2022.
Line 100. I would delete the "support" here.
Line 125. Did you just stack the Sentine-1 data from both "ascending" and "descending" orbit together? This will lead to distortions over mountain areas if you do so.
Line 195. I know these metrices make sense, but you may explain why they were chosen.
Line 199. Could you explain why temporal metrices were only derived from Sentinel-1? There are much more Sentinel-2 images (Figure 2).
Table 2."Number" ->"Dimension".
Line 205. Could you show us the distribution of training samples?
Line 233. How this threshold (SAD > 0.125) was determined?
Line 235. Why a pixel can be regarded as water if its MNDWI is greater than 0.12? It's not a robust method.
Line 240. Without seeing your training samples, I presume many pixels identified as changed may not change from NIS to IS if they are far from IS. Could you do a quick ablation experiment to better demonstrate the effectiveness of the sample migration?
Line 265. The methods (and figures) you described here are similar to that in Li et al., (2015). You should at least cite it.
Li, X., Gong, P., and Liang, L.: A 30-year (1984–2013) record of annual urban dynamics of Beijing City derived from Landsat data, Remote Sens. Environ., 166, 78–90, https://doi.org/10.1016/j.rse.2015.06.007, 2015.
Line 272. If your results follow the assumption that transition from IS to NIS is rare, you may reduce the results to a single band where pixel indicates the time when IS was first detected, instead of putting annual data into separate bands.
Figure 6. It seems that most of your IS test samples locate in or near cities. Could you provide the ISA density around the IS samples? Courtesy of higher spatial resolution, buildings and roads in rural regions can be better delineated in Sentinel images. Therefore, accuracy assessment over rural regions for 10-m IS mapping is important.
Line 302. Are there new (or interesting) findings from your results?
Line 328. "0:05" ->"0.05"?
Table 5. It's would be more interesting to compare your results with GHS and GISA-10m (I mentioned above). They are both 10-m thematic mappings, same as you did here. I strongly suggest you to do so.
Line 377. In fact, according to your results (Table 5), the ESA-2020 achieved higher PA in all cities, indicating that ESA-2020 has less omissions. This is contradictory with the statement here that ESA-2020 ignores buildings in rural areas. Could you explain it?
Citation: https://doi.org/10.5194/essd-2022-251-RC1 -
CC1: 'Reply on RC1', Aizhu Zhang, 17 Sep 2022
The manuscript "High-resolution and Multitemporal Impervious Surface Mapping in the Lancang-Mekong Basin with Google Earth Engine" describes a new IS map for Mekong basin. To this aim, the authors applied visual-interpreted samples to multi-source features derived from Sentinel-1/2 images. The whole framework is routine and some key parameters should be further clarified or discussed. Current accuracy assessment showed better performance than some datasets. However, comparison with the state-of-art datasets is missing, the superiority to the existing works remains unclear. Moreover, the test samples seem to be located mostly in urban and sub-urban areas, leading to incomplete accuracy assessment over rural regions. The structural of manuscript is clear but the writing should be carefully revised. As such, I would suggest a rejection.
Line 70. Many 10-m global datasets have been released, such as the GHS-BUILT-S (Corbane et al., 2021) from EC JRC and the GISA-10m (Huang et al., 2022) from Wuhan University.
Corbane, C., Syrris, V., Sabo, F., Politis, P., Melchiorri, M., Pesaresi, M., Soille, P., and Kemper, T.: Convolutional neural networks for global human settlements mapping from Sentinel-2 satellite imagery, Neural Comput. Appl., 33, 6697–6720, https://doi.org/10.1007/s00521-020-05449-7, 2021.
Huang, X., Yang, J., Wang, W., and Liu, Z.: Mapping 10 m global impervious surface area (GISA-10m) using multi-source geospatial data, Earth Syst. Sci. Data, 14, 3649–3672, https://doi.org/10.5194/essd-14-3649-2022, 2022.
Response: Thanks for your suggestions. We noted the release of the global 10m impervious surface products, but (1) the globally shared samples and methods may not yield good results for a special region like the Lancang-Mekong Basin, and (2) these products are only available in one phase and do not meet the requirements of dynamic analysis. For this reason, we conducted the development of a 10m multi-temporal impervious surface for the Lancang-Mekong Basin.
Line 100. I would delete the "support" here.
Response: Thanks, we will take this comment.
Line 125. Did you just stack the Sentine-1 data from both "ascending" and "descending" orbit together? This will lead to distortions over mountain areas if you do so.
Response: Our experimental results show that after combining the mask of DEM data, the direct stacking of the two does not have much effect on the misclassification of the mountains.
Line 195. I know these metrices make sense, but you may explain why they were chosen.
Response: The 7*7 size was obtained based on existing studies and experiments.
Line 199. Could you explain why temporal metrices were only derived from Sentinel-1? There are much more Sentinel-2 images (Figure 2).
Response: The original Sentinel-2 images are more than the Sentinel-1 images, but there are many clouds and rain in the Lancang-Mekong Basin, resulting in poor quality of some optical images. After we eliminate this part, the number of Sentinel-2 images is not enough to calculate the temporal features.
Table 2."Number" ->"Dimension".
Response: Thanks, we will take this comment.
Line 205. Could you show us the distribution of training samples?
Response: The samples were obtained by random sampling and manually divided into sample polygons of the same type around them.
Line 233. How this threshold (SAD > 0.125) was determined?
Response: We experimentally obtained the division thresholds by selecting a number of sample points, including both changing and non-changing classes.
Line 235. Why a pixel can be regarded as water if its MNDWI is greater than 0.12? It's not a robust method.
Response: Thanks, this is indeed not a robust approach, but we still use it in order to reduce the computational effort.
Line 240. Without seeing your training samples, I presume many pixels identified as changed may not change from NIS to IS if they are far from IS. Could you do a quick ablation experiment to better demonstrate the effectiveness of the sample migration?
Response: The ablation experiment at large-scale is rather complicated, and we can only give a comparison of the precision in some areas. Without sample migration, the overall precision of the results in 2021 will decrease from 91.45% to 89.86%.
Line 265. The methods (and figures) you described here are similar to that in Li et al., (2015). You should at least cite it.
Li, X., Gong, P., and Liang, L.: A 30-year (1984–2013) record of annual urban dynamics of Beijing City derived from Landsat data, Remote Sens. Environ., 166, 78–90, https://doi.org/10.1016/j.rse.2015.06.007, 2015.
Response: Thanks, we should cite it here.
Line 272. If your results follow the assumption that transition from IS to NIS is rare, you may reduce the results to a single band where pixel indicates the time when IS was first detected, instead of putting annual data into separate bands.
Response: Thanks, this is a good way to reduce the dataset.
Figure 6. It seems that most of your IS test samples locate in or near cities. Could you provide the ISA density around the IS samples? Courtesy of higher spatial resolution, buildings and roads in rural regions can be better delineated in Sentinel images. Therefore, accuracy assessment over rural regions for 10-m IS mapping is important.
Response: We obtained the validation samples by stratified sampling, and although the samples are less at rural areas, as a whole, the representativeness is sufficient. We will use a detailed comparison to illustrate the effect in rural areas.
Line 302. Are there new (or interesting) findings from your results?
Response: I am sorry to say that the main purpose of this paper is to propose mapping methods for large scale impervious surfaces in the Lancang-Mekong Basin and to develop a usable product. But no effort was spent on analyzing the new (or interesting) findings.
Line 328. "0:05" ->"0.05"?
Response: Thanks, we have a spelling error here.
Table 5. It's would be more interesting to compare your results with GHS and GISA-10m (I mentioned above). They are both 10-m thematic mappings, same as you did here. I strongly suggest you to do so.
Response: We are trying to compare with these two datasets, but that will take some time. At least for now, visually, we have a better product than GHS, which is too fragmented in the Lancang-Mekong Basin.
Line 377. In fact, according to your results (Table 5), the ESA-2020 achieved higher PA in all cities, indicating that ESA-2020 has less omissions. This is contradictory with the statement here that ESA-2020 ignores buildings in rural areas. Could you explain it?
Response: As you mentioned before, the results of the accuracy validation are strongly related to the distribution of the samples, and the stratified samples are mostly scattered in urban areas, which is the reason why the ESA accuracy is higher than ours. However, in terms of visual effects, ESA has a certain omission in rural areas.
Citation: https://doi.org/10.5194/essd-2022-251-CC1
-
CC1: 'Reply on RC1', Aizhu Zhang, 17 Sep 2022
-
RC2: 'Comment on essd-2022-251', Anonymous Referee #2, 04 Sep 2022
This manuscript provided 10-m resolution impervious surface area data using Sentinel-1 and Sentinel-2 data in Lancang-Mekong Basin. Overall, the novelty of this paper is not qualified for publication in ESSD for the following reasons.
(1) Mapping of impervious surface area at the global scale with 10-m or 30-m resolutions has been widely reported in many studies. The method employed in this paper doesn’t show significant improvement regarding the novelty of mapping approaches as well as the mapped results and influences.
(2) Although the migration strategy of training samples across years highlights in this paper, some concerns may introduce uncertainties in this study. For example, the sampling in 2016 may omit those urbanized regions in 2021, which may bias the classification results (see Table 4). Also, the threshold of SAD is challenging to determine if these pixels have experienced change or not.
(3) The temporal consistency check is not the novelty of this work. This approach was initially proposed in the RSE paper in 2015 (in the case of Beijing). Unfortunately, I didn’t find the citation of this work when they introduced the “temporal consistency check” part, and the overall framework is highly similar to that one.
Title: A 30-year (1984–2013) record of annual urban dynamics of Beijing City derived from Landsat data.
(4) It is unfair to compare the local mapping results with those of global products because it is a space-specific application at the river basin. I think the proposed approach cannot be extended directly to other regions (even the globe).
Citation: https://doi.org/10.5194/essd-2022-251-RC2 -
CC2: 'Reply on RC2', Aizhu Zhang, 17 Sep 2022
This manuscript provided 10-m resolution impervious surface area data using Sentinel-1 and Sentinel-2 data in Lancang-Mekong Basin. Overall, the novelty of this paper is not qualified for publication in ESSD for the following reasons.
(1) Mapping of impervious surface area at the global scale with 10-m or 30-m resolutions has been widely reported in many studies. The method employed in this paper doesn’t show significant improvement regarding the novelty of mapping approaches as well as the mapped results and influences.
Response: In this paper, we propose a framework for fusing multi-source data based on the GEE platform to address the lack of available multi-temporal 10m impervious surface datasets due to multiple factors in the Lancang-Mekong Basin, and finally publish a dataset of impervious surfaces in the Lancang-Mekong Basin for the period 2016-2021.
(2) Although the migration strategy of training samples across years highlights in this paper, some concerns may introduce uncertainties in this study. For example, the sampling in 2016 may omit those urbanized regions in 2021, which may bias the classification results (see Table 4). Also, the threshold of SAD is challenging to determine if these pixels have experienced change or not.
Response: In fact, the rate of renewal of impervious surface categories is much less than the rate of area expansion. That is, there will be the same material and spatial characteristics of impervious surfaces in urbanized areas in 2021 that existed in 2016. Considering this, the ultimate purpose of samples migration is simply to avoid the error that occurs in imaging over many years of images, and if this error does not exist, it is in fact only necessary to use the model trained in 2016 for classification. About the threshold of SAD, we experimentally obtained the division thresholds by selecting a number of sample points, including both changing and non-changing classes.
(3) The temporal consistency check is not the novelty of this work. This approach was initially proposed in the RSE paper in 2015 (in the case of Beijing). Unfortunately, I didn’t find the citation of this work when they introduced the “temporal consistency check” part, and the overall framework is highly similar to that one.
Title: A 30-year (1984–2013) record of annual urban dynamics of Beijing City derived from Landsat data.
Response: Thanks, we should cite it here.
(4) It is unfair to compare the local mapping results with those of global products because it is a space-specific application at the river basin. I think the proposed approach cannot be extended directly to other regions (even the globe).
Response: It is true that this comparison is unfair and that the globally shared samples and methods may not yield good results for a special region like the Lancang-Mekong Basin. However, this is the only method used for comparison, as no other products specific to the LMC Basin exist. It also illustrates the need to develop products for a particular region, which is the point of our proposed methodology.
Citation: https://doi.org/10.5194/essd-2022-251-CC2
-
CC2: 'Reply on RC2', Aizhu Zhang, 17 Sep 2022
-
RC3: 'Comment on essd-2022-251', Christopher Brown, 20 Sep 2022
The authors present a method for extrapolating annualized annotations for binary impervious surface (IS) / non-impervious surface (NIS) into future years when NIS -> IS transitions are assumed to be insignificantly probable. The authors also present a method for converting a short sequence of annual IS / NIS binary classifications into a sequence congruent with the previous assumption regarding transitions. The authors combine these methods in a standard supervised Sentinel-1 / Sentinel-2 random forest classification workflow in Google Earth Engine to produce annual IS / NIS maps of the Lancang-Mekong basin from 2016 to 2021. These maps are compared to a number of existing global LULC products with an IS class, or a class the authors considered comparable, using a dataset created specifically for this purpose. The results of the author's assessment indicate good overall agreement with their sample.
I would begin by asking the authors to carefully copy-review their manuscript: I find that there are grammatical and/or punctuation errors on nearly all of the proof pages. In particular I suggest paying special attention to extraneous usage of "the" before nouns, and general comma usage. I also suggest replacing verbs used in a subjective context such as "excellent" or "state-of-the-art" with the less strongly worded "popular" or "modern." There are also opportunities to unify language e.g. "dynamic monitoring of IS" is more commonly known as "change monitoring," and the authors are not consistent in their wording. I find the structure of the manuscript's narrative generally well paced and direct.
The author's characterization of the Sentinel-1 and Sentinel-2 missions was imprecise. Line 62 asserts Sentinel-1 has a 10m "spatial" resolution, however Sentinel-1's IW spatial resolution is actually 5m x 20m with the GRD product specifically resampled to a pixel resolution of 10m (https://sentinels.copernicus.eu/web/sentinel/missions/sentinel-1/overview/mission-summary). Line 103 discusses Sentinel-2 "image blocks," which is not a Sentinel-2 data product I'm aware of, and lines 103-104 discuss Sentinel-1 and SRTM "scenes," which are not part of the ARD scheme used in delivering either of these products. On line 123, the authors state Sentinel-1's revisit to be 6 days, and while this is true of the instrument, the availability of the data used by the authors (GRD IW) is highly variable between orbits, location, and polarization (https://www.researchgate.net/figure/Sentinel-1-constellation-observation-revisit-and-coverage-scenario-April-2021-10_fig2_351632665).
I have concerns with the expert system introduced by the authors to perform sample migration, that which formed the basis of their strategy to extend their training data. Besides a lack of ablation analysis on the many parameters, Table 4 (which would benefit from NIS column totals) raises doubts about the author's procedure. Considering that a fundamental assumption in the author's framework is that IS -> NIS changes are not permitted, one would expect points migrated to successive years to decrease with time. This is not the case, and e.g. points that were not migrated to 2017 (15,649 total) are in fact migrated to 2018 (16,075), calling into question the validity of this approach since it violates the IS -> NIS assumption between 2017-2018. At the extremes, it would be possible for a point to exist exclusively in 2016 and 2021. I am also concerned with the use of pixel-counting statistics used in Equations 2-5, as it is not stated whether or not the authors account for processing baseline updates in the underlying Sentinel-2 collection. Failing to do so would lead to double-counting and so erroneous results.
I appreciate that the procedural diagrams in Figures 3, 4, and 5 were easy to follow, which is somewhat of a rarity.
I also have concerns with the temporal consistency check introduced by the authors. From first principles, the authors are, in my opinion, suggesting a rather complicated procedure to post-process a length 6 vector with only 64 possible values, 57 of which are actually invalid. The authors do not state the boundary considerations for the sliding window, which is important as it effectively defines the IS change sequence semantics in ambiguous cases like "101010." Furthermore, that such ambiguous sequences are collapsed to "valid" sequences seems like a fallacious assumption in the overall methodology when a "transition" or "low-confidence" classification may be more appropriate.
Little information is given regarding the stratification performed to collect validation data.
I find that Figures 10 and 11 and the corresponding section offer little insight or value into the author's data.
The author's compute area statistics from their maps yet do not offer any confidence intervals or uncertainties. These statistics are all but the bare minimum required to publish estimations of area based on remotely sensed data.
I recommend the authors discuss the potential for change dynamics to effect their agreement statistics with the external data products outside of the sampling year. The authors conclude on Lines 348-349 that similar accuracies to WorldCover-10 2020 in 2021 indicate that the sample migration strategy was successful, but since no baseline was established, it is unclear whether similar performance can be attributable to a combination of class-shift and definition penalizing WorldCover-10, extra training data introduced in re-using points across years, or the nuances of the algorithm introduced by the authors.
One of the biggest potential contributions from this effort would have been the underlying training data sampled from 2016, and I encourage the authors to make this available.
While there is good work in this research effort, I believe the authors should intensely proofread their writing and try to keep their language exact and consistent. I also believe the novelty, the suggested sample migration and temporal smoothing, demands more methodological consideration and emphasis in the manuscript. The algorithms introduced also require more justification to contextualize their effects on the accuracy assessment. The authors should seek existing literature on area estimation to ensure confidences are published with the computed statistics. My overall recommendation is for reconsideration after major revisions.
Citation: https://doi.org/10.5194/essd-2022-251-RC3
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
805 | 282 | 52 | 1,139 | 52 | 62 |
- HTML: 805
- PDF: 282
- XML: 52
- Total: 1,139
- BibTeX: 52
- EndNote: 62
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1