A Framework to Retrieve Water Quality Parameters in Small, Optically Diverse Freshwater Ecosystems Using Sentinel-2 MSI Imagery

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. Satellite remote sensing has the advantage of rapid data acquisition over large areas and plays an important role in the field of ecological environment monitoring of lakes, reservoirs, and other water bodies. In the introduction section, it is recommended to supplement references to the latest published relevant literature.

2. The content of the comparison of accuracy, advantages and disadvantages analysis between this method and traditional remote sensing inversion algorithms for lake water quality parameters should be supplemented.

3. Are there any limitations to the research area for the methods proposed in this paper? Can they be extended to more regions? If they can be extended, what are the requirements for the basic data of the research area? It is recommended to supplement these contents in the discussion section.

4. What are the main innovations of this paper reflected in? It is suggested to refine them in the last paragraph of the introduction section.

Author Response

Response to Reviewer 1

 

1. Satellite remote sensing has the advantage of rapid data acquisition over large areas and plays an important role in the field of ecological environment monitoring of lakes, reservoirs, and other water bodies. In the introduction section, it is recommended to supplement references to the latest published relevant literature.

Thank for this recommendation. We added the following recent citations to the introduction section:

Alves e Santos, D.R., Martinez, J.M., Olivetti, D., Zumak, A., Guimarães, D., Aniceto, K., Severo, E., Ferreira, O., Harmel, T., Cordeiro, M., Fillizola, N., Sell, B., Fernandes, D., Souto, C., Roig, H., 2024. Sentinel-2 MSI image time series reveal hydrological and geomorphological control of the sedimentation processes in an Amazonian hydropower dam. International Journal of Applied Earth Observation and Geoinformation 128, 103786. doi:10.1016/j.jag.2024.

Diehl, R.M., Underwood, K.L., Watt, R., Hamshaw, S.D., Pahlevan, N., 2024. Evaluating opportunities for broad-scale remote sensing of total suspended solids on small rivers. Remote Sensing Applications: Society and Environment 35, 101234. doi:10.1016/j.rsase.2024.

Liu, X., Steele, C., Simis, S., Warren, M., Tyler, A., Spyrakos, E., Selmes, N., Hunter, P., 2021. Retrieval of chlorophyll-a concentration and associated product uncertainty in optically diverse lakes and reservoirs. Remote Sensing of Environment 267, 112710. doi:10.1016/j.rse.2021.

Wang, S., Li, J., Zhang, W., Cao, C., Zhang, F., Shen, Q., ... and Zhang, B. 2021. A dataset of remote-sensed Forel-Ule Index for global inland waters during 2000–2018. Scientific Data, 8(1), 26.

 

2. The content of the comparison of accuracy, advantages and disadvantages analysis between this method and traditional remote sensing inversion algorithms for lake water quality parameters should be supplemented.

We agree. The use of traditional remote sensing inversion algorithms has been discussed before, such as in the works of Spyrakos et al. (2018), Neil et al. (2019) and Pahlevan et al. (2021). These works have shown that it is impossible to retrieve water quality parameters in large scale (i.e. local to regional to continental scales) using a single calibrated algorithm due to the different concentrations of optically active water constituents and assumptions of the inversion algorithms. We added a sentence about this topic to the introduction section:

“In addition, research has also shown the limitation of retrieving water quality parameters at regional scale using a single calibrated algorithm for each parameter due to the different concentrations of optically active water constituents and the assumptions and simplifications of the inversion algorithms, and the potential of pre-selecting algorithms based on OWT classification, making it possible to automatically select the best models in each case based only on the water spectral response (Spyrakos et al., 2018; Neil et al., 2021; Liu et al., 2021).”

 

3. Are there any limitations to the research area for the methods proposed in this paper? Can they be extended to more regions? If they can be extended, what are the requirements for the basic data of the research area? It is recommended to supplement these contents in the discussion section.

We did discuss about the limitations of the methods in the Discussion, sections 4.2 and 4.3. They are related, and thus dependant, to the adjacency effect, so environments where it is stronger (such as urban lakes) will have less accurate results, and on the reflectance level, with low Rrs lakes also possibly having less accuracy due to a lack of a chl-a model specific for dark water lakes. We also have much less data on tropical regions, thus a larger validation effort is need for such areas.

And yes, this method was tested on this study area, but it was proposed to be generic enough to be extrapolated to other regions. The data required are Rrs data, if available, and water quality data to validate the retrieval of the parameters. We complemented this to the Discussion (last paragraph):

“Future work will further validate our framework (both Rrs and the water quality parameters) in smaller lakes (∼1 ha) and in other regions, such as in tropical and subtropical areas, where we expect, for instance, variability in the performance of the atmospheric correction method due to variability in the concentration of gases and aerosols, and of the chl-a algorithms due to increased light availability and chl-a concentrations, and the packaging effect [65]. We also recommend further studies to validate Rrs retrievals from satellite imagery over lakes with diverse areas and possibly in rivers, where remote sensing retrievals are even more challenging. We also recommend, when possible, to conduct samplings over transects, to evaluate how distance to land impacts the adjacency effect and to assess its impacts on the retrieval of parameters.”

 

4. What are the main innovations of this paper reflected in? It is suggested to refine them in the last paragraph of the introduction section.

Thank you for this suggestion. The innovations of this paper are reflected in the study of the adjacency effects over small lakes, which is original to our knowledge, on the capacity of the Sentinel-2 MSI sensors to consistently retrieve water quality parameters in a regional scale using Optical Water Types, with the models not being tuned specifically for one or a certain number of lakes, and the accuracy and consistency of retrieving turbidity using Optical Water Types, which has not been assessed before. We updated the last paragraph of the introduction section with these innovations described in the form of objectives and research questions:

“Therefore, this work aims at validating a framework for retrieving two water quality parameters, chlorophyll-a concentration and turbidity, from small water bodies with variable characteristics located in the Garonne and Adour river basins in Southwest France. Specifically, we: a) evaluate the accuracy of remote sensing reflectance derived from Sentinel-2 MSI in small lakes, and the possible impacts of adjacency effects on these retrievals; b) assess the performance of inversion models based on optical water types (OWTs) in retrieving the water quality parameters across lakes with diverse water optical characteristics; and c) demonstrate the consistency of the method in capturing the spatial and temporal variability of these two parameters. The research questions we aim to answer are: a) how adjacency effects alters Rrs in small lakes, and does it hinder the retrieval of water quality parameters in these cases?; b) can the Sentinel-2 MSI sensors consistently retrieve water quality parameters in small lakes in a regional scale, without the models being tuned specifically for one or a certain number of lakes?; and c) what is the accuracy and consistency of retrieving turbidity using Optical Water Types?”

Reviewer 2 Report

Comments and Suggestions for Authors

This study investigates small inland water bodies by developing and validating a Sentinel-2 MSI-based framework for retrieving chlorophyll-a and turbidity in lakes as small as 3 hectares. The key innovations include: (1) implementation of a GRS specifically optimized for small water bodies, (2) comprehensive validation across 108 lakes with diverse optical properties. My detailed recommendations are as follows:
1. While the manuscript emphasizes spatial resolution as key for small lakes, it should more explicitly discuss other fundamental differences from large lake remote sensing.
2. The removal of pixels with negative Rrs in red bands is noted as a quality control step. However, if there exist negative Rrs in real application, how to process the pixel?
3. In figure 8, the chl-a pattern around Lake Laragou appears artificially sharp. Please confirm whether in situ data support this spatial heterogeneity and analyze the reason.
4. The chl-a retrieval uses 13 OWTs while turbidity uses only 4. Was there evidence that turbidity requires fewer classes, or is this an artifact of available training data? 
5. WaterDetect is applied per-scene. For time series analysis (Figs. 11-12), how was consistency maintained given potential water level changes? Were manual corrections needed for small lakes where area fluctuations could affect pixel inclusion?
7. The workflow combines multiple error sources (GRS correction, OWT classification, algorithm selection). Is there a quantitative estimate of how these compound in the final products? 
8. Given the complexity of the multi-step approach, were end-to-end deep learning methods (e.g., CNN trained on in situ matchups) explored? If not, please justify the preference for physical/empirical hybrid methods.

Author Response

Response to Reviewer 2

 

This study investigates small inland water bodies by developing and validating a Sentinel-2 MSI-based framework for retrieving chlorophyll-a and turbidity in lakes as small as 3 hectares. The key innovations include: (1) implementation of a GRS specifically optimized for small water bodies, (2) comprehensive validation across 108 lakes with diverse optical properties. My detailed recommendations are as follows:

1. While the manuscript emphasizes spatial resolution as key for small lakes, it should more explicitly discuss other fundamental differences from large lake remote sensing.

We thank you for your pertinent comments. Yes, there are other differences between the remote sensing of large and small lakes, namely: the adjacency effect, which was extensively discussed in the manuscript since its intensity is a function of distance from land, and thus depends on the size (or shape) of the lakes; the increased possibility of bottom effect since many small lakes are optically shallow, and this effect is a result of the mixing of the water signal with the bottom substrate, affecting all estimations of water quality parameters; and the detection of water pixels, since the proportion of land-water mixed pixels is much higher than that of large lakes. The latter specifically also justify the use of WaterDetect in our work, a robust, image-based algorithm to detect water surfaces.

We added the following sentence to the 3^rd paragraph of the introduction to describe these other differences:

“Other challenges in the remote sensing of small lakes are the increased possibility of bottom effect, when the water signal is mixed with the substrate at the bottom of the lake, as a result of shallow depth, and the detection of water-only pixels, since the proportion of land-water mixed pixels is much higher than that of large lakes [24].”

 

2. The removal of pixels with negative Rrs in red bands is noted as a quality control step. However, if there exist negative Rrs in real application, how to process the pixel?

We are not sure if we understood well your commentary. It is physically impossible to have negative Rrs in reality, but when processing atmospheric correction over inland waters, it is possible to find negative Rrs as a result of over-correction, for example due to aerosol scattering or sunglint reflectance. In these cases, the negative Rrs impedes the application of the inversion models, therefore it is recommended to remove all pixels with negative Rrs from the processing to avoid wrong results. We selected the red band since the water Rrs in this spectrum is usually lower than in the blue or green bands (thus it is more common to find negative Rrs in the red spectrum), and because it is used in the majority of the inversion models applied in this work.

 

3. In figure 8, the chl-a pattern around Lake Laragou appears artificially sharp. Please confirm whether in situ data support this spatial heterogeneity and analyze the reason.

We are not sure how to reply to this question, as it is a bit ambiguous. Lake Laragou, letter (e) in the map, has a very homogenous superficial chl-a. The green layer around the lake is actually the visible lake colour taken from the Sentinel-2 image, as the chl-a maps are superimposed on this image. It is worth remembering that we use WaterDetect for detecting water surface pixels, and we also use an inward buffer on the polygon to remove mixed land-water pixels, resulting in this visible difference in all 5 lakes shown in this image. We added this information to the captions of the figure of spatial maps.

As we were unsure about you remark, maybe you mean Lake Bocage, letter (a) in the map, where we do have spatialised field data that supports this chl-a pattern, with points B1, B2, B3 and B10 (map in Figure 1) having consistently higher chl-a. The reasons could be the wind, which can blow in the southeast direction and drive phytoplankton in this direction, especially in the case of an algae bloom which occurs in this cyanobacteria-dominated lake in early autumn and can be the case on this date. Another reason could be the turbulence of the water, with this lake being used for water sports like jet ski and that are concentrated in the northern part of the lake, increasing water turbulence and reducing phytoplankton productivity. The last and less likely possibility could be the inputs from the small agricultural field close to the southern part of the lake that are drained to the lake in the event of rain, slowly increasing productivity in this area and creating the observed chl-a gradient. We did not include this discussion to the manuscript since understading the dynamics of the lakes is not our objective in particular, but we can include it if requested, while we hope that it answers your question and demonstrates the robustness of our approach.

 

4. The chl-a retrieval uses 13 OWTs while turbidity uses only 4. Was there evidence that turbidity requires fewer classes, or is this an artifact of available training data?

There is evidence that less classes are needed for turbidity estimates, as quite often it can be roughly estimated using the Rrs peak from the green to NIR bands (B3 to B8A), depending on the OWT. Two of the works which base our study, Jiang et al. (2021 and 2023), showed that 4 is the optimal number of classes to retrieve TSS with its quasi-analytical algorithm, which is applied in our method in a slighly modified way, using the 4 OWTs that were derived independently by Cordeiro (2022).

 

5. WaterDetect is applied per-scene. For time series analysis (Figs. 11-12), how was consistency maintained given potential water level changes? Were manual corrections needed for small lakes where area fluctuations could affect pixel inclusion?

For the time series analysis, we applied a criteria to post-process each image for each lake, establishing a minimum of 9 pixels and at least 20% of all possible pixels in each lake, for each scene. Therefore, in cases where lakes could dry out or almost dry, if less than 20% of pixel were available, no water quality was produced for such dates.

Regarding WaterDetect, we did not do any manual corrections as the algorithm is applied per scene and is unsupervised and independant of previous runs. If some pixels were dry, we expect it to not detect them as water, as shown in our tests and in previous validations of this algorithm (Peña-Luque et al., 2021). Some mixed water-land pixels could also be detected as water by the algorithm, but these should be mask out by the inward buffer and the post-processing filters we applied.

 

6. The workflow combines multiple error sources (GRS correction, OWT classification, algorithm selection). Is there a quantitative estimate of how these compound in the final products?

Unfortunately we could not produce any quantitative estimate of the compound errors in our study, as we did not produce any water quality parameter using in situ Rrs data, since the majority of the data collected (by the water agency and in the gravel pit lakes) consists of only water quality data. We showed that GRS has an error varying from 16% to 32% for bands 1 to 5, while the error for chl-a was 56% and for turbidity was 47%. It is hard to estimate the propagation of the errors solely from these results, as we do not know how much of this error comes from the OWT classification and how much comes from the models.

 

7. Given the complexity of the multi-step approach, were end-to-end deep learning methods (e.g., CNN trained on in situ matchups) explored? If not, please justify the preference for physical/empirical hybrid methods.

It is an interesting suggestion. Although we understand that these are powerful tools, we did not test those methods for the same reason why we did not try to calibrate the physical or empirical methods: to maintain the generalisation potential of our methodology, so that it can be tested and applied in other regions. Locally calibrating algorithms to our study area would considerably reduce their ability to be general.

Reviewer 3 Report

Comments and Suggestions for Authors

1. The paper states that Sentinel-2 MSI is the first sensor capable of continuously monitoring water quality parameters in small water bodies; however, it does not clearly articulate the specific innovations compared to previous studies.

2. The manuscript does not provide sufficient details regarding the representativeness of the selected dataset in terms of geography, climate, and land use types.

3. Remote sensing technology has made significant progress not only in inland water quality monitoring (e.g., rivers and lakes), but also in the broader domain of ocean remote sensing. It is recommended that the authors include a review of related studies in the introduction, such as underwater image captioning based on the fusion of visual and textual information and Underwater Image Captioning with AquaSketch-Enhanced Cross-Scale Information Fusion.

4. The paper employs the WaterDetect algorithm for water body detection, but it is recommended that the authors elaborate on the rationale behind selecting this algorithm and provide a comparative analysis with other available methods.

5. The manuscript briefly mentions the field data collection process but lacks detailed explanation on how data accuracy and consistency were ensured.

6. Although the paper highlights the applicability of the proposed method to small lakes, it does not thoroughly discuss its potential applicability to other types of water bodies.

Author Response

Response to Reviewer 3

 

1. The paper states that Sentinel-2 MSI is the first sensor capable of continuously monitoring water quality parameters in small water bodies; however, it does not clearly articulate the specific innovations compared to previous studies.

Thank you for your suggestion. We summarised our innovations in the Conclusion section: “This study is the first to use remote sensing data to accurately retrieve water quality parameters from small lakes (<1000 ha) at a regional scale. [...] mapping long-term water quality from water bodies as small as 3 ha. To achieve this, we employed a framework for generating per-pixel chlorophyll-a concentration and turbidity maps, combining two different sets of OWTs and different algorithms for each parameter. We validated the methods with a large dataset of small lakes in Southern France, demonstrating the robustness of the framework.”

“The Rrs retrieved using GRS, correcting for both atmospheric and sunglint effects,

was found to be consistent in our case studies. However, we observed artefacts caused by adjacent land, particularly a strong signal in the NIR spectrum. While this did not hinder the retrieval of parameters due to the filters employed to detect contaminated pixels, correcting for this effect remains a priority for improving product quality.”

 

2. The manuscript does not provide sufficient details regarding the representativeness of the selected dataset in terms of geography, climate, and land use types.

This is clearly described in the Methods, section 2.1. Study area and in situ data. We did not find pertinent to add more information pertinent to the lakes individually, as it is possible to see in Figure 1, the dataset is distributed all over the area described in this section:

“The study area is located in Southwest France, within the Garonne and Adour river basins (Figure 1). This region is considered one of the most vulnerable in Europe in terms of water resources due to a drier climate and frequent drought episodes, which have been intensifying with climate change [6]. To mitigate this, numerous water reservoirs have been built to store part of the runoff during the wet season in order to face the dry season, to ensure water supply for drinking and irrigation. However, very few of these reservoirs are monitored, limiting the comprehension of water quality dynamics and the impacts of droughts on their ecosystem health.

The Garonne river is France’s third longest river, originating in the Spanish Pyrenees mountains and discharging into the Atlantic Ocean at the Gironde Estuary near Bordeaux. It has a drainage area of approximately 84,000 km² which includes major tributaries such as the Ariège, Lot, and Tarn rivers. Geographically, the Garonne basin is divided into three main regions: the plains, surrounded by the Pyrenees to the south and the Massif Central to the north-east. The Adour river, a smaller catchment, drains an area of about 17,000 km², and also originates in the Pyrenees and flows into the Atlantic Ocean. Both river basins host numerous dams that store water for the dry season, which typically lasts from summer to early autumn. The climate in these watersheds is composed of Mediterranean climate on the Mediterranean coast, a continental type in the south, and an oceanic climate along the Atlantic coast. Rainfall is more concentrated between May and June, with an average annual precipitation of approximately 900 mm. Land use in these regions is predominantly agricultural, particularly in the plains, while forests are more common around the Massif Central.”

 

3. Remote sensing technology has made significant progress not only in inland water quality monitoring (e.g., rivers and lakes), but also in the broader domain of ocean remote sensing. It is recommended that the authors include a review of related studies in the introduction, such as underwater image captioning based on the fusion of visual and textual information and Underwater Image Captioning with AquaSketch-Enhanced Cross-Scale Information Fusion.

We found this to be an unusual request. Our work focuses on inland waters surface remote sensing, therefore we did not use tools from ocean remote sensing or underwater imaging.

 

4. The paper employs the WaterDetect algorithm for water body detection, but it is recommended that the authors elaborate on the rationale behind selecting this algorithm and provide a comparative analysis with other available methods.

The rationale behind selecting WaterDetect is describe in the Methods, section 2.3. Processing Sentinel-2 MSI L1C images:

“This algorithm has been tested in other studies [33], showing high efficiency in detecting even small water bodies, being much more consistent than applying only spectral indices such MNDWI, for example.”

There are different methods available for detecting water in satellite imagery, but WaterDetect was shown to be much more efficient than other methods in the comparative analysis already done by Peña-Luque et al. (2021), being superior even when compared to techniques using radar. It also offers compatibility with MAJA imagery, which has a very robust cloud masking algorithm (later described in this same paragraph), and is also robust for running individual images as well as time series.

 

5. The manuscript briefly mentions the field data collection process but lacks detailed explanation on how data accuracy and consistency were ensured.

We are sorry but we did not understand your question. All field work followed the conventional methods of measuring environmental variables and were consistent, as described in the Methods, section 2.2. Field campaigns for validation and assessment of adjacency effect.

 

6. Although the paper highlights the applicability of the proposed method to small lakes, it does not thoroughly discuss its potential applicability to other types of water bodies.

Thank you for this recommendation. We validated our approach for lakes up to ~6500 ha, therefore considering the large range of size found in the validation dataset, we expect our approach to also work for larger lakes. However, a nice suggestion is to also test our approach in lotic environments, where conditions can be even more challenging due to lower width (lower number of pixels), increased possibility of bottom effect due to water shallowness and transparency, and increased sunglint due to increased turbulence. We added this possibility at the end of the Discussion section:

“We also recommend further studies to validate Rrs retrievals from satellite imagery over lakes with diverse areas and shapes and possibly in rivers, where remote sensing retrievals are even more challenging.”

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I have no more suggestions at present.

Reviewer 2 Report

Comments and Suggestions for Authors

No comments.

Reviewer 3 Report

Comments and Suggestions for Authors

Well revised.