the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 12 Jan 2026
Optimizing SAR Flood Extent Mapping in Central Chile: The Critical Role of Image Timing
Abstract. This study critically quantifies the temporal uncertainty inherent in flood extent estimation using Sentinel-1 SAR data in the high-relief, flash-flood-prone river systems of Central Chile, following the extreme events of 2023. We applied the iterative Jaccard optimization framework to five sites in the Maule and Mataquito watersheds, identifying the Difference Image Index (DII) as the most robust flood indicator. Our key finding is that the estimation of maximum flood extent is fundamentally limited by the timing of the SAR acquisition. River gauge analysis confirmed a flash-flood regime with an extremely rapid recession rate (river height dropping ∼ 50 % within four days of the peak). This rapid drainage means that a delay of 24–48 hours results in a severe underestimation of the true flood footprint. While the DII performed best, overall Jaccard scores remained low (≤ 0.6). We conclude that the method's accuracy is primarily constrained by physical limitations- namely, the rapid recession rate and complex topography- rather than the calibration technique itself. Relying solely on the Sentinel 1 revisit cycle is insufficient for operational mapping in such dynamic environments, and we recommend integrating SAR monitoring with hydraulic modeling or high-frequency aerial surveys to accurately interpolate the maximum flood extent.
- Preprint
(20636 KB) - Metadata XML
- BibTeX
- EndNote
Status: closed (peer review stopped)
- RC1: 'Comment on egusphere-2025-6409', Anonymous Referee #1, 06 Feb 2026
-
RC2: 'Comment on egusphere-2025-6409', Anonymous Referee #2, 25 Mar 2026
I read the manuscript with great interest and appreciate that the authors investigated aspects of SAR-based flood mapping that go beyond image classification accuracy. Furthermore, the study presents an interesting method using optical measurements to optimize the threshold applied to SAR indices. However, several issues need to be addressed before publication. Generally, the study needs to be better integrated into the scientific literature, and the differences between optical and SAR-based flood maps need to be more thoroughly considered.
My detailed feedback can be found here:
- Introduction: You never explain how water/floods can be detected from SAR observations or what methods are commonly used to do so. Please better integrate your method/study into the existing scientific literature!
- Line 19: You mention active remote sensors and then discuss implications of the backscatter signal. I suggest you properly introduce radar/SAR sensors to make your arguments more comprehensible.
- Line 21: When referring to a "channel," do you mean "polarization" of a SAR sensor?
- Line 26: It is inconsistent to state that the indices are established without citing any publications that make use of them.
- Line 34: You repeatedly use the term "robust framework." What do you mean by this, and what makes your framework robust?
- Figure 1: Why not use different colour bars for different variables (maximum daily rain and accumulated rain)?
- Figure 2: Consider using a distinct colour for the five analysed areas to make them more visible in contrast to the drainage network and gauging stations.
- Line 79: I assume the input to your pre-processing is Sentinel-1 GRD data. Did you obtain this data directly from Copernicus or Google Earth Engine (GEE)?
- Line 83: You argue that the value of the applied threshold is commonly used, yet you cite no supporting publications.
- Line 84: You use the optical-based Global Surface Water dataset for masking SAR-based flood maps. Do you anticipate any implications from this approach? This should be addressed when discussing the limitations.
- Line 104: Using flood maps from optical satellite data is not equivalent to "ground truth"; it is a reference dataset with its own limitations. Furthermore, you do not explain how the optical flood maps are retrieved from the Sentinel-2 MNDWI data.
- Line 137: Since the Jaccard index is an essential component of your method, it would be beneficial to describe it in more detail, particularly its interpretation.
- Figure 7: This figure clearly lacks a legend. What do blue and black represent in this context?
- Figure 8: Adding a legend would also be beneficial here. What do the green lines signify?
- Line 189: Since you are already using optical data to optimize your SAR-based flood maps, the appropriate method for evaluating the performance of your approach would be to compare the results against a third independent dataset. As the main goal of your study is to evaluate the critical role of image timing rather than classification accuracy, you should be cautious with performance-related statements in this section. This should also be noted in the study limitations.
- Line 205: Referencing the table in the appendix, which shows the temporal differences between SAR and optical observations would be beneficial. Additionally, you do not mention that optical and SAR data would never observe the exact same water extent due to the specific characteristics of their observation methods. Is this not also a limitation of the approach?
- Table A1: Using optical data observed multiple days apart from the corresponding SAR observation seems questionable. In such cases, is the approach applicable?
- Table A1: When discussing the differences between ascending and descending observations, you frequently use the term "orbit," when what you actually mean is "orbit direction." Please correct this!
I hope these comments are helpful to improve the manuscript and I see this study as a valuable contribution to SAR-based flood mapping.
Citation: https://doi.org/10.5194/egusphere-2025-6409-RC2
Status: closed (peer review stopped)
-
RC1: 'Comment on egusphere-2025-6409', Anonymous Referee #1, 06 Feb 2026
The manuscript aims to quantify the uncertainty in maximum flood extent retrieval from satellite data induced by the time gap between flood peak and satellite image acquisition. For this purpose, Sentinel-1 SAR data acquired during two flash flood events over five locations in Central Chile were analysed by training a classifier based on three different image-based indices against a reference flood map derived from optical satellite imagery. The results are interpreted as an indication that the flood extent retrieved from Sentinel-1 imagery acquired several days after the rainfall peak did not reflect the maximum flood extent during the event.
The topic of uncertainty in flood extent mapping due to image timing is of high importance to both the remote sensing and the hydrological community as flood extents are highly dynamic and, especially in the case of flash flood events, may undergo changes at time scales below the typical image acquisition interval of satellite missions. Considering the vast amount of studies dedicated to flood mapping algorithms, this topic has received relatively attention than it deserves. However, I think that the study contains a number of flaws, which I would like to ask the authors to take into account.
Specific comments:
- Pages 1-2, lines 13-41: The current state of the art regarding the impacts of SAR image (or satellite image, in general) timing on the retrieval of maximum flood extent is not discussed in detail. While this topic has been somewhat overlooked, several studies have addressed this topic in recent years (e.g. Dasgupta et al., 2021; García-Pintado et al., 2013; Gobeyn et al., 2017; Notti et al., 2018; Zhao et al., 2025).
- Flood mapping methodology:
- Pages 5-7: There seem to be some discrepancies in the description of the computation of the SAR-based indices (DII, RI, NDFI) in equations 1-3 in the manuscript and the original definitions by Hamidi et al. (2023). While $mean(\sigma_0_{reference})$ is defined as the mean value “across the entire reference image raster” in this manuscript, i.e. computed over all pixels of a single image, Hamidi et al. (2023) define it as the pixel-wise mean across a pre-event image stack, i.e. computed along the time dimension of a number of images. This has important implications: if $mean(\sigma_0_{reference})$ is a single value, the shape of the image histogram will not change when computing the indices, only an offset value is subtracted. As a consequence, the indices will not provide any measure of change with respect to non-flood backscatter values.
- If an actual change detection using a pre-flood reference image (or image stack) is performed there should be no need to use a reference water occurrence layer as permanent water should be included already in $mean(\sigma_0_{reference})$.
- Page 5, lines 94-95: I wonder why there was such a long time gap between the Sentinel-1 reference images and the flood events. The reference images were from 2019, while the events occurred in 2023. In the meantime, land-cover changes may have occurred which could lead to alterations in backscatter, e.g. asphalt that could be mistaken for water surfaces.
- Page 7, lines 102-108: It is unclear how the authors arrived at a binary reference flood map from Sentinel-2 MNDWI for the optimisation. A threshold should be applied here as well. Often a value of MNDWI = 0 is applied but this might be non-optimal, e.g. in the case of mixed pixels. This part is also not clearly described in Hamidi et al. (2023) but, in my opinion, deserves some attention.
- Page 9, lines 136-139: More unbiased estimates of the classification accuracy would be obtained by retaining 20-25% of the training samples for validation, ideally using a spatial resampling scheme for cross-validation.
- Page 14, lines 174-176: As the authors rightly point out, the threshold optimisation is likely to be impacted by the time lag between Sentinel-1 and Sentinel-2 image acquisitions and a bias is introduced which contributes to the total error induced now not only by the acquisition timing of Sentinel-1 but also Sentinel-2. Would it be possible to separately quantify the contribution of this error?
- I also wonder how practical this approach is for flood mapping in a real-world scenario because an optical image acquired during the flood event is required for the threshold optimisation. During the flood peak, the chance of obtaining a cloud-free image of the flooded areas is likely to be lower than of getting a SAR acquisition which is unaffected by cloud cover. In this manuscript, the Sentinel-2 image was typically acquired one day after the Sentinel-1 image, when cloud cover was probably already lower than at the time of the flood peak.
- It is not clear which SAR polarisation the authors used. While it seems likely that VV was used, this is not stated in the text.
- Quantification of temporal uncertainty:
- Ideally, the uncertainty caused by the timing of the image acquisition during the descending limb of the flood hydrograph could be quantified by comparing the true flood extent at the time of peak flood with the SAR-derived flood extent a few days after. However, no such ground truth was available for the two events studied in this case and the results of the flood extent mapping were rather interpreted in the light of the time that had passed between the precipitation peak and the image acquisition. Hence, the uncertainty was not “critically quantified” as stated in the objectives of the study, but a more qualitative assessment of the likely uncertainty is instead provided.
- I wonder if a proxy for a ground truth of maximum flood extent could be extracted from post-event multispectral imagery by comparing it to pre-event data. Especially if the flood resulted in damage to vegetation or deposition of sediment the flooded areas may be visible for several weeks.
- The number of two events over a relatively limited geographic area is small for extrapolating the conclusions to flash floods in general.
- A more robust approach would be to test the approach with events characterised by different combinations of flood recession rates and image acquisition timings.
Further comments:
Page 1, line 2: It is not clear to the reader at this point what “the extreme events of 2023” are.
Page 1, line 20: Active sensors typically do not have higher temporal resolution than passive sensors. Regarding free access, the only currently source of freely accessible SAR data is Sentinel-1 (and NISAR in the near future), while others, such as TerraSAR-X, CosmoSky-Med or ICEYE, have restricted data access.
Page 4, lines 59-60: I wonder how the complex terrain of the study region influences the precipitation estimates by GPM IMERG. Different studies have shown biases in IMERG precipitation over mountainous regions (Bulovic et al., 2020; Rojas et al., 2021; Sharifi et al., 2019). For completeness, which version of GPM IMERG was used?
Figure 2: It is very hard to see the locations of the runoff gauges due to the strong colours of the satellite image basemap. The source of the image should be provided in the caption. However, the map would be easier to understand if only topography was used as a base map, e.g. a (possibly hillshaded) DEM or OpenTopoMap.
Page 5, lines 82-84: Unfortunately, the Sentinel-1 data hosted on Google Earth Engine do not have layover/radar shadow masks. These masks (Kropatsch & Strobl, 1990) can, however, be computed with open source software packages if the Sentinel-1 scenes are downloaded. In that case, no further topographical filtering would be necessary. What is the advantage of using Google Earth Engine in comparison to processing the data offline, other than computational reasons?
Page 9, line 145: “specifically the absence of Sentinel-1A data”. I do not understand how the three SAR-based indices were derived then, as all three require a reference image.
Page 10, line 159: While it is likely that the estimated flood extent is lower than the flood extent at the time of flood peak, I think it cannot be stated here that the there is a significant reduction if the real maximum flood extent during the event is unknown.
Figure 5: The dashed green line is hardly visible.
Figure 6: It would be easier for the reader to see the differences in performance between the indices if the same scaling of the y axis was kept.
Page 16, line 250: Both hydraulic modelling and high-frequency aerial surveys carry a high cost. I wonder if such efforts are practicable, especially in remote regions where no high-resolution DTM or suitable aircraft may be available. Also, what would be the impact of using data from commercial SAR constellations, such as ICEYE or Capella, given the fact that some of their data may be freely available for emergency situations?
Table A1: I find this table very helpful for the interpretation of the results as it provides the exact timings of the image acquisitions with respect to the peak of the event and I think it would be better placed and discussed in the main part rather than the appendix. The table should also contain the dates of the reference image(s).
References:
Bulovic, N., McIntyre, N., & Johnson, F. (2020). Evaluation of imerg v05b 30-min rainfall estimates over the high-elevation tropical andes mountains. Journal of Hydrometeorology, 21(12), 2875-2892.
Dasgupta, A., Hostache, R., Ramsankaran, RAAJ, Schumann, G.J-P., Grimaldi, S., Pauwels, V.R.N., Walker, J.P. (2021) On the Impacts of Observation Location, Timing, and Frequency on Flood Extent Assimilation Performance. Water Resources Research, 57(2), e2020WR028238. https://doi.org/10.1029/2020WR028238
García-Pintado, J., Neal, J. C., Mason, D. C., Dance, S. L., & Bates, P. D. (2013). Scheduling satellite-based SAR acquisition for sequential assimilation of water level observations into flood modelling. Journal of Hydrology, 495, 252–266. https://doi.org/10.1016/j.jhydrol.2013.03.050
Gobeyn, S., Van Wesemael, A., Neal, J., Lievens, H, Van Eerdenbrugh, K., De Vleeschouwer, N., Vernieuwe, H., Schumann, G., Di Baldassarre, G., De Baets, B., Bates, P.D., Verhoest, N.E.C. (2017) Impact of the timing of a SAR image acquisition on the calibration of a flood inundation model. Advances in Water Resources, 100, 126-138. https://doi.org/10.1016/j.advwatres.2016.12.005
Hamidi, E., Peter, B.G., Munoz, D.F., Moftakhari, H., Moradkhani, H. (2023). Fast Flood Extent Monitoring Wuth SAR Change Detection Using Google Earth Engine. IEEE Transactions on Geoscience and Remote Sensing, 61, 4201419
Kropatsch, W., and D. Strobl (1990), The Generation of SAR layover and shadow maps from digital elevation models, IEEE Trans. Geosci. Remote Sens., 28 (1), 98–107; doi:10.1109/36.45752.
Notti, D., Giordan, D., Caló, F., Pepe, A., Zucca, F., & Galve, J. P. (2018). Potential and Limitations of Open Satellite Data for Flood Mapping. Remote Sensing, 10(11), 1673. https://doi.org/10.3390/rs10111673
Rojas, Y., Minder, J. R., Campbell, L. S., Massmann, A., & Garreaud, R. (2021). Assessment of GPM IMERG satellite precipitation estimation and its dependence on microphysical rain regimes over the mountains of south-central Chile. Atmospheric Research, 253, 105454.
Sharifi, E., Eitzinger, J., & Dorigo, W. (2019). Performance of the state-of-the-art gridded precipitation products over mountainous terrain: A regional study over Austria. Remote Sensing, 11(17), 2018.
Zhao, J., Li, M., Li, Y., Matgen, P., Chini, M. (2025). Urban Flood Mapping Using Satellite Synthetic Aperture Radar Data: A review of characteristics, approaches, and datasets. IEEE Geoscience and Remote Sensing Magazine, 13(1), 237-268. https://doi.org/10.1109/MGRS.2024.3496075
Citation: https://doi.org/10.5194/egusphere-2025-6409-RC1 -
RC2: 'Comment on egusphere-2025-6409', Anonymous Referee #2, 25 Mar 2026
I read the manuscript with great interest and appreciate that the authors investigated aspects of SAR-based flood mapping that go beyond image classification accuracy. Furthermore, the study presents an interesting method using optical measurements to optimize the threshold applied to SAR indices. However, several issues need to be addressed before publication. Generally, the study needs to be better integrated into the scientific literature, and the differences between optical and SAR-based flood maps need to be more thoroughly considered.
My detailed feedback can be found here:
- Introduction: You never explain how water/floods can be detected from SAR observations or what methods are commonly used to do so. Please better integrate your method/study into the existing scientific literature!
- Line 19: You mention active remote sensors and then discuss implications of the backscatter signal. I suggest you properly introduce radar/SAR sensors to make your arguments more comprehensible.
- Line 21: When referring to a "channel," do you mean "polarization" of a SAR sensor?
- Line 26: It is inconsistent to state that the indices are established without citing any publications that make use of them.
- Line 34: You repeatedly use the term "robust framework." What do you mean by this, and what makes your framework robust?
- Figure 1: Why not use different colour bars for different variables (maximum daily rain and accumulated rain)?
- Figure 2: Consider using a distinct colour for the five analysed areas to make them more visible in contrast to the drainage network and gauging stations.
- Line 79: I assume the input to your pre-processing is Sentinel-1 GRD data. Did you obtain this data directly from Copernicus or Google Earth Engine (GEE)?
- Line 83: You argue that the value of the applied threshold is commonly used, yet you cite no supporting publications.
- Line 84: You use the optical-based Global Surface Water dataset for masking SAR-based flood maps. Do you anticipate any implications from this approach? This should be addressed when discussing the limitations.
- Line 104: Using flood maps from optical satellite data is not equivalent to "ground truth"; it is a reference dataset with its own limitations. Furthermore, you do not explain how the optical flood maps are retrieved from the Sentinel-2 MNDWI data.
- Line 137: Since the Jaccard index is an essential component of your method, it would be beneficial to describe it in more detail, particularly its interpretation.
- Figure 7: This figure clearly lacks a legend. What do blue and black represent in this context?
- Figure 8: Adding a legend would also be beneficial here. What do the green lines signify?
- Line 189: Since you are already using optical data to optimize your SAR-based flood maps, the appropriate method for evaluating the performance of your approach would be to compare the results against a third independent dataset. As the main goal of your study is to evaluate the critical role of image timing rather than classification accuracy, you should be cautious with performance-related statements in this section. This should also be noted in the study limitations.
- Line 205: Referencing the table in the appendix, which shows the temporal differences between SAR and optical observations would be beneficial. Additionally, you do not mention that optical and SAR data would never observe the exact same water extent due to the specific characteristics of their observation methods. Is this not also a limitation of the approach?
- Table A1: Using optical data observed multiple days apart from the corresponding SAR observation seems questionable. In such cases, is the approach applicable?
- Table A1: When discussing the differences between ascending and descending observations, you frequently use the term "orbit," when what you actually mean is "orbit direction." Please correct this!
I hope these comments are helpful to improve the manuscript and I see this study as a valuable contribution to SAR-based flood mapping.
Citation: https://doi.org/10.5194/egusphere-2025-6409-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,230 | 593 | 121 | 1,944 | 130 | 172 |
- HTML: 1,230
- PDF: 593
- XML: 121
- Total: 1,944
- BibTeX: 130
- EndNote: 172
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript aims to quantify the uncertainty in maximum flood extent retrieval from satellite data induced by the time gap between flood peak and satellite image acquisition. For this purpose, Sentinel-1 SAR data acquired during two flash flood events over five locations in Central Chile were analysed by training a classifier based on three different image-based indices against a reference flood map derived from optical satellite imagery. The results are interpreted as an indication that the flood extent retrieved from Sentinel-1 imagery acquired several days after the rainfall peak did not reflect the maximum flood extent during the event.
The topic of uncertainty in flood extent mapping due to image timing is of high importance to both the remote sensing and the hydrological community as flood extents are highly dynamic and, especially in the case of flash flood events, may undergo changes at time scales below the typical image acquisition interval of satellite missions. Considering the vast amount of studies dedicated to flood mapping algorithms, this topic has received relatively attention than it deserves. However, I think that the study contains a number of flaws, which I would like to ask the authors to take into account.
Specific comments:
Further comments:
Page 1, line 2: It is not clear to the reader at this point what “the extreme events of 2023” are.
Page 1, line 20: Active sensors typically do not have higher temporal resolution than passive sensors. Regarding free access, the only currently source of freely accessible SAR data is Sentinel-1 (and NISAR in the near future), while others, such as TerraSAR-X, CosmoSky-Med or ICEYE, have restricted data access.
Page 4, lines 59-60: I wonder how the complex terrain of the study region influences the precipitation estimates by GPM IMERG. Different studies have shown biases in IMERG precipitation over mountainous regions (Bulovic et al., 2020; Rojas et al., 2021; Sharifi et al., 2019). For completeness, which version of GPM IMERG was used?
Figure 2: It is very hard to see the locations of the runoff gauges due to the strong colours of the satellite image basemap. The source of the image should be provided in the caption. However, the map would be easier to understand if only topography was used as a base map, e.g. a (possibly hillshaded) DEM or OpenTopoMap.
Page 5, lines 82-84: Unfortunately, the Sentinel-1 data hosted on Google Earth Engine do not have layover/radar shadow masks. These masks (Kropatsch & Strobl, 1990) can, however, be computed with open source software packages if the Sentinel-1 scenes are downloaded. In that case, no further topographical filtering would be necessary. What is the advantage of using Google Earth Engine in comparison to processing the data offline, other than computational reasons?
Page 9, line 145: “specifically the absence of Sentinel-1A data”. I do not understand how the three SAR-based indices were derived then, as all three require a reference image.
Page 10, line 159: While it is likely that the estimated flood extent is lower than the flood extent at the time of flood peak, I think it cannot be stated here that the there is a significant reduction if the real maximum flood extent during the event is unknown.
Figure 5: The dashed green line is hardly visible.
Figure 6: It would be easier for the reader to see the differences in performance between the indices if the same scaling of the y axis was kept.
Page 16, line 250: Both hydraulic modelling and high-frequency aerial surveys carry a high cost. I wonder if such efforts are practicable, especially in remote regions where no high-resolution DTM or suitable aircraft may be available. Also, what would be the impact of using data from commercial SAR constellations, such as ICEYE or Capella, given the fact that some of their data may be freely available for emergency situations?
Table A1: I find this table very helpful for the interpretation of the results as it provides the exact timings of the image acquisitions with respect to the peak of the event and I think it would be better placed and discussed in the main part rather than the appendix. The table should also contain the dates of the reference image(s).
References:
Bulovic, N., McIntyre, N., & Johnson, F. (2020). Evaluation of imerg v05b 30-min rainfall estimates over the high-elevation tropical andes mountains. Journal of Hydrometeorology, 21(12), 2875-2892.
Dasgupta, A., Hostache, R., Ramsankaran, RAAJ, Schumann, G.J-P., Grimaldi, S., Pauwels, V.R.N., Walker, J.P. (2021) On the Impacts of Observation Location, Timing, and Frequency on Flood Extent Assimilation Performance. Water Resources Research, 57(2), e2020WR028238. https://doi.org/10.1029/2020WR028238
García-Pintado, J., Neal, J. C., Mason, D. C., Dance, S. L., & Bates, P. D. (2013). Scheduling satellite-based SAR acquisition for sequential assimilation of water level observations into flood modelling. Journal of Hydrology, 495, 252–266. https://doi.org/10.1016/j.jhydrol.2013.03.050
Gobeyn, S., Van Wesemael, A., Neal, J., Lievens, H, Van Eerdenbrugh, K., De Vleeschouwer, N., Vernieuwe, H., Schumann, G., Di Baldassarre, G., De Baets, B., Bates, P.D., Verhoest, N.E.C. (2017) Impact of the timing of a SAR image acquisition on the calibration of a flood inundation model. Advances in Water Resources, 100, 126-138. https://doi.org/10.1016/j.advwatres.2016.12.005
Hamidi, E., Peter, B.G., Munoz, D.F., Moftakhari, H., Moradkhani, H. (2023). Fast Flood Extent Monitoring Wuth SAR Change Detection Using Google Earth Engine. IEEE Transactions on Geoscience and Remote Sensing, 61, 4201419
Kropatsch, W., and D. Strobl (1990), The Generation of SAR layover and shadow maps from digital elevation models, IEEE Trans. Geosci. Remote Sens., 28 (1), 98–107; doi:10.1109/36.45752.
Notti, D., Giordan, D., Caló, F., Pepe, A., Zucca, F., & Galve, J. P. (2018). Potential and Limitations of Open Satellite Data for Flood Mapping. Remote Sensing, 10(11), 1673. https://doi.org/10.3390/rs10111673
Rojas, Y., Minder, J. R., Campbell, L. S., Massmann, A., & Garreaud, R. (2021). Assessment of GPM IMERG satellite precipitation estimation and its dependence on microphysical rain regimes over the mountains of south-central Chile. Atmospheric Research, 253, 105454.
Sharifi, E., Eitzinger, J., & Dorigo, W. (2019). Performance of the state-of-the-art gridded precipitation products over mountainous terrain: A regional study over Austria. Remote Sensing, 11(17), 2018.
Zhao, J., Li, M., Li, Y., Matgen, P., Chini, M. (2025). Urban Flood Mapping Using Satellite Synthetic Aperture Radar Data: A review of characteristics, approaches, and datasets. IEEE Geoscience and Remote Sensing Magazine, 13(1), 237-268. https://doi.org/10.1109/MGRS.2024.3496075