A Multi-Sensor Approach to Characterize Winter Water-Level Drawdown Patterns in Lakes

Round 1

Reviewer 1 Report

The paper is interesting and the method you propose can be used in other areas.

I would suggest some improvements in the text.

1/ Some explanation on Sentinel -1 VV  backscattering differences from water. land, vegetation would be useful.

2/ Please explain which Sentinel-1 product you use. GRDH or SLC ?

3/ Table 1 - please check  Sentinel -1 spatial resolution. 10 m is too high. 

4/ Line 205  sentence "divided one-year 204 (April to March) in-situ" is not clear

5/ In the introduction it would be useful to describe the hydrological regime of the study area. You mention only winter and annual precipitation totals. For lakes and their water management operation more important is the river runoff changes in the seasons. 

Author Response

Reviewer #1

Comments and Suggestions for Authors:

The paper is interesting and the method you propose can be used in other areas. I would suggest some improvements in the text.

Comment:  Some explanation on Sentinel -1 VV backscattering differences from water. land, vegetation would be useful.

Response: We appreciate the reviewer highlighting the need for further explanation of the Sentinel-1 backscatter characteristics of different land cover types in sentinel-1 SAR data. SAR sensors like Sentinel-1 emit microwave pulses and measure the intensity of backscattered energy returned to the sensor. The strength of the backscatter signal depends on surface properties like roughness, moisture content and structure (Meyer, 2019; Esmaeili et al., 2023).

For open water bodies like lakes, the smooth surface acts as a specular reflector that reflects most of the incident radar pulses away from the sensor, resulting in very low backscatter and dark tone in SAR imagery (Santoro and Wegmuller, 2014). In contrast, vegetation, urban areas and bare land have higher surface roughness leading to diffuse scattering of radar pulses in multiple directions, including back to the sensor (Rozenstein et al., 2016). This results in higher backscatter intensity and brighter tone compared to water. Specifically, vegetation contains discontinuities between leaves, branches and stem as well as water content that enhance scattering. Built infrastructure in urban areas like buildings and roads acts as corner reflector that strongly bounce signals back. Bare land and soils exhibit variable moisture content and surface roughness that modulates backscatter intensity.

In this study, we utilized the low backscatter response from lake water surfaces compared to surrounding land cover types to effectively discriminate lakes in the Sentinel-1 scenes. By applying thresholding techniques like the Otsu algorithm to the backscatter images, we were able to classify pixels as either water or non-water. Overall, the distinct backscatter characteristics of different land cover types enabled reliable mapping of lake water areas indicative of water level variations using Sentinel-1 time series analysis. We agree that further details on the backscatter response from different land covers would help support the methods. Therefore, we have added some of the above explanations in the revised manuscript (Section 2.2 (Lines: 165-174)).

Comment:  Please explain which Sentinel-1 product you use. GRDH or SLC?

Response: We have used the Sentinel-1 GRD product. We have added the detail in the revised version of the manuscript (Table 1; Section 2.2 (Line: 150)).

Comment: Table 1 - please check Sentinel -1 spatial resolution. 10 m is too high.

Response:  We have used Sentinel-1 GRD data available within Google Earth Engine. There are three spatial resolutions (10, 25, 40 m) at which data are available for this product within GEE. The highest spatial resolution of the available data is 10 m as mentioned in the data description within the GEE. We have updated Table 1 and included multiple spatial resolutions for S1-SAR data in the revised manuscript. We have also clarified in the revised manuscript that we resampled S1-SAR data at 30 m spatial resolution to have consistency among the three sensors used in this study (Lines: 155-159).

Comment:  Line 205 sentence "divided one-year (April to March) in-situ" is not clear

Response: We have added additional explanation on how Carmignani et al. (2021) divided each annual cycle (April to March) data into two seasons: summer (April to September) and winter (October to March) to classify winter drawdown lakes. The revised sentences in the manuscript are as follow: “To differentiate WD lakes from ND lakes, Carmignani et al. (2021) divided each annual cycle (April to March) of in-situ water level data into summer (April to September) and winter (October to March) periods. They found that WD lakes exhibited significantly lower water levels during the winter compared to summer, while ND lakes showed the opposite seasonal pattern” (Section 2.5 (Lines: 248-252)).

Comment:  In the introduction it would be useful to describe the hydrological regime of the study area. You mention only winter and annual precipitation totals. For lakes and their water management operation more important is the river runoff changes in the seasons.

Response: Good suggestion. We have added a new “Study Area” subsection (Section 2.1) at the beginning of the Method section and described the hydrological regime of the study area as per your suggestion. The revised version of the manuscript has the following sentences under the study area subsection: “The study focused on lakes in Massachusetts, located in the northeastern United States. This region experiences a humid continental climate with warm, wet summers and cold, snowy winters (Wilkin 2006). Precipitation in the region is generally evenly distributed throughout the year except for the wet season, with an annual average rainfall of 1016-1270 mm (Siddique et al., 2020). River flows in the region are typically characterized by higher flows during leaf-off periods (~November through April) and spring snowmelt, and lowest flows during summer months (Tu, 2009; Demaria et al., 2016). Natural lakes in Massachusetts fill during the high spring flows and gradually decline during the drier summer-fall months as evapotranspiration exceeds precipitation inputs (McHorney and Neill, 2007; Shuman and Donnelly, 2006). Most WD lakes implement drawdowns in the fall by releasing water downstream when river flows are lowest. Past research indicates the timing of these drawdowns, which coincide with the region’s seasonal precipitation and temperature patterns, can significantly impact physical habitat structure and macrophytes in Massachusetts lakes (Carmignani and Roy, 2021)” (Section 2.1 (Lines: 111-125)).

Reviewer 2 Report

The manuscript proposed a multi-sensor approach to characterize winter water level drawdown patterns in lakes. The manuscript is generally well prepared with some interesting results. However, the introduction and method sections should be improved as suggested below.

Specific comments:

1.     Introduction: The methods of estimating the winter water level drawdown from optical and SAR data should be carefully reviewed. Moreover, the contribution should be highlighted as the novelty in method is marginal.

2.     Line119-120: The preprocessing was made before achieved on the GEE. The authors did not make such preprocessing steps.

3.     A flowchart is required for section 2 to provide an overview of how the results were achieved.

4.     Section 2.3: It’s not clear how the accuracy was evaluated.

5.     Line 211-212: To my knowledge, KNN is an unsupervised method without requiring a training process.

6.     Line 212-213: Not clear why daily surface water extents were created as the KNN. The potential uncertainty should be addressed

7.     Line 217-230: The cross-validation process is not clear. Did the authors mean 7-fold cross-validation?

8.     An independent data section is required.

Author Response

Reviewer #2

Comments and Suggestions for Authors:

The manuscript proposed a multi-sensor approach to characterize winter water level drawdown patterns in lakes. The manuscript is generally well prepared with some interesting results. However, the introduction and method sections should be improved as suggested below.

Specific comments:

Comment:  Introduction: The methods of estimating the winter water level drawdown from optical and SAR data should be carefully reviewed. Moreover, the contribution should be highlighted as the novelty in method is marginal.

Response: We have thoroughly reviewed the methods for estimating winter water level drawdowns using optical and SAR data. The approach leverages established methods for water classification from satellite imagery, including thresholding methods like Otsu for optical data and backscatter thresholding for SAR data. However, we applied and validated these techniques specifically for detecting winter water level changes in lakes, which poses unique challenges due to factors like ice cover. Through our accuracy assessment using data from in-situ loggers, we identified key limitations of optical sensors related to ice and sun glint effects and determined that SAR provided the most reliable water level estimates for winter conditions. We have alreay highlighted the novelty of our study in the response to the Academic Editor and in the revised version of the manuscript Methods (Sections 2.5 (Lines: 297-301); 2.6 (Lines: 304-306); 2.7 (Lines: 333-336)) and Discussion sections (Sections 4.1 (Lines: 543-546); 4.4 (Lines: 698-701)).

While this research built upon established remote sensing methods, we believe it makes incremental but meaningful progress in addressing the challenges of mapping lake hydrologic regimes during winter conditions. The spatial and temporal scope of our analysis (166 lakes over 5 years throughout Massachusetts) would not be feasible without satellite remote sensing capabilities and cloud computing tools leveraged. Our study helps lay the groundwork for improved mapping of understudied lake management practices like winter water level drawdown across broad geographic regions in the future. The revised manuscript provides more details and clearly articulates the novel contribution of our study.

Comment:   Line119-120: The preprocessing was made before achieved on the GEE. The authors did not make such preprocessing steps.

Response: We apologize for the lack of clarity regarding the preprocessing steps conducted prior to implementing our methods in Google Earth Engine. In the revised manuscript, we have clarified that radiometric calibration and orthorectification of Sentinel 1 SAR data were already conducted before archiving the data on GEE (Lines: 150-152). However, we performed additional preprocessing steps such as resampling all three sensors’ data to a common spatial resolution (30 m) and speckle filtering of Sentinel 1-SAR data prior to implementing the Otsu thresholding (Lines: 154-158). 

Comment:  A flowchart is required for section 2 to provide an overview of how the results were achieved.

Response: Thank you for the suggestion to add a flowchart of the overall study methodology. We agree this would significantly improve the clarity and visualization of the multi-step workflow applied in this analysis. In the revised manuscript, we have added a flowchart as per reviewer’s suggestion (Figure 2; Lines: 130-137).

Comment:  Section 2.3: It’s not clear how the accuracy was evaluated.

Response: In the revised manuscript, we have expanded the description of how accuracy was evaluated when comparing the satellite derived surface water area to the in-situ water level data. Specifically, we first found the total number of matched dates between the available satellite images and in-situ logger data during the study period for each of the 21 lakes. We then qualitatively examined the alignment in the patterns and trends between the satellite-derived surface water area time series and the in-situ water level time series. We looked for consistent increases and decreases that aligned between the two datasets, indicative of accurate capture of water level fluctuations. Additionally, we described how factors like lake size, shape, and surrounding terrain affected the accuracy of surface area as a proxy for water level changes. We have added these details on the accuracy assessment methods to Section 2.4 (Lines: 214-219) to provide a clearer explanation.

Comment:  Line 211-212: To my knowledge, KNN is an unsupervised method without requiring a training process.

Response: After double checking references, we confirmed K-nearest neighbors is actually considered a supervised machine learning approach and requires a training process. Therefore, KNN is not an unsupervised method.

Comment:  Line 212-213: Not clear why daily surface water extents were created as the KNN. The potential uncertainty should be addressed.

Response: Thank you for highlighting the need to justify and evaluate uncertainty from creating daily surface water extent estimates from the ~6 days satellite observation frequency. This was an important step for preparing the time series data for the KNN classification model which uses tslearn package. In the revised manuscript, we have explained that the KNN algorithm requires continuous time series data as input. Therefore, we used linear interpolation between the actual satellite observation dates to generate daily surface water extent values. This allowed us to create a consistent time step series across all lakes for the KNN model implementation. While interpolation introduces some uncertainty on days without true observations, we chose linear interpolation as a simple method to create a continuous dataset while minimizing assumption of non-linear changes between observations. Importantly, the ~6 days revisit frequency still captured the major seasonal fluctuations in water levels critical for distinguishing winter drawdown. We have added a paragraph to include above clarification in the Method section (Section 2.5; Lines:  258-266).  We have also acknowledged the potential uncertainty from interpolation in the limitations and included consideration of non-linear changes between observation dates as an area for future refinement (Section 4.2; Lines: 617-620).

Comment: Line 217-230: The cross-validation process is not clear. Did the authors mean 7-fold cross-validation?

Response: We apologize for the ambiguity. In the revised manuscript, we have clarified the cross-validation methodology by specifying use of 7-fold cross-validation when evaluating the classification model performance. In the 7-fold cross-validation, we divided the test dataset of 17 total lakes into 7 groups, each containing 3 lakes (2 WD and 1 ND lake). For each fold, 3 lakes held out for model testing while the remaining 14 lakes were used for model training. This process was repeated 7 times so each lake was included in testing. The accuracy results were then averaged across the 7 folds. We have updated the text to state 7-fold cross-validation was implemented and specified the details of the training and test splits used in each fold iteration (Section 2.5; Lines: 270, 276-278, 280).

Comment:  An independent data section is required.

Response: We agree that clearly outlining the data sources and access details will improve the reproducibility and transparency of this work. In the revised manuscript, we have added a new section titled “Data Availability” at the end of the manuscript text before the References. This section provides details on how to access the different datasets used in this study such as: “The satellite data from Landsat, Sentinel-1, and Sentinel-2 are available via public archives from USGS and Copernicus. These satellite data can also be accessed using the Google Earth Engine platform. Precipitation data can be accessed using NASA Giovanni web-based application interface. The in-situ lake level data used in this study are available in ScienceBase (https://www.sciencebase.gov/catalog/item/64b1760fd34e70357a2a0133). Lake bathymetry data can be accessed via Mass Wildlife (https://www.mass.gov/orgs/division-of-fisheries-and-wildlife).  The lake boundary shapefile used for spatial analysis is from US Fish and Wildlife Service National Wetland Inventory, available at (https://www.fws.gov/program/national-wetlands-inventory/download-state-wetlands-data). Satellite data and codes used in this study are available in a public repository at (https://gitlab.com/gee_codes/winter-drawdown)” (Lines: 732-741).

Reviewer 3 Report

Manuscript: remotesensing- 2841614

General comments: The manuscript titled ‘A Multi-Sensor Approach to Characterize Winter Water Level 2 Drawdown Patterns in Lakes’ is a study to characterize water drawdown in lake water levels for the state of Massachusetts in the United States. The study describes the evaluation of satellite data in tracking water levels with in-site data and follows with mapping the exposed water area for all the lakes in Massachusetts. The manuscript is well written and describes the mapping techniques in detail along with validating it with in-situ data. I just have a few minor comments for the authors to address. 

Comment #1: 

Figure 1 is not cited in the manuscript text. 

Comment #2:

Line 248: Can you please provide an explanation in the manuscript text in a few sentences on why the time series were not deseasonalized and detrended before feeding into kNN algorithm? 

Comment #3:

Line 346: For Figure 4, there was a significant decline in water area in in-situ data whereas Sentinel-1 SAR didn’t capture the additional water drawdown? What is a potential reason for that?

Comment #4:

Line 355-358: Could you please elaborate on the physical reasons for why Sentinel-1 is showing the water area returned to normal before the in-site water levels. Is this consistent across different years and across different lakes?

Comment #5: 

Line 371-372: This comment again brings the question of not deseasonalizing and detrending the time series. Have the authors tried to compare the results of 12-months and 6-months WD and ND on deseasonalized and detrended time series and compared it with 3-months results? Since in 3-month there is no inherent seasonality but for 12-month there is going to be strong interannual seasonality and a weaker but still present seasonality for 6-month time series.

Comment #6:

Line 526-528: Since machine learning algorithms are sensitive to unequal sample sizes, the accuracy results provided in the manuscript would be biased in favor of larger sample size. I understand the limitation of availability of in-situ data, but I would suggest the authors to provide a discussion on how it would affect the results in identifying WD vs ND lakes using kNN. You can add a few more sentences in section 4.2 at the end of the first paragraph discussing this limitation in context of your present results.

Author Response

Reviewer #3

 General comments: The manuscript titled ‘A Multi-Sensor Approach to Characterize Winter Water Level Drawdown Patterns in Lakes’ is a study to characterize water drawdown in lake water levels for the state of Massachusetts in the United States. The study describes the evaluation of satellite data in tracking water levels with in-site data and follows with mapping the exposed water area for all the lakes in Massachusetts. The manuscript is well written and describes the mapping techniques in detail along with validating it with in-situ data. I just have a few minor comments for the authors to address.

Comment: Figure 1 is not cited in the manuscript text.

 Response: Thank you for catching the mistake. We have cited Figure 1 in the revised manuscript (Section 2.1 (Line: 112)).

Comment:  Line 248: Can you please provide an explanation in the manuscript text in a few sentences on why the time series were not deseasonalized and detrended before feeding into kNN algorithm?

Response: Thank you for raising this point. We wanted to retain the intrinsic seasonality and trends that differentiate drawdown and non-drawdown hydrologic regimes. Specifically, the kNN algorithm relies on detecting similarities in the overall shape and pattern of time series data to classify them. Winter drawdown lakes exhibit a distinct seasonal decline in water levels during fall-winter months, while non-drawdown lakes do not show this sharp downward trend. Removing the seasonal component through deseasonalization may eliminate this key differentiation in hydrologic regimes that the kNN uses for classification. Likewise, not detrending retains the directionality and magnitude of water level changes that distinguish drawdown versus non-drawdown lakes.Since drawdown lakes exhibit a distinct and consistent declining trend in water levels not seen in non-drawdown lakes, detrending may eliminate this signal. However, we agree that exploring pre-processing as an option to potentially improve classification is worthwhile. We have acknowledged this as an area for future refinement of the methodology in the revised manuscript (Section 2.6; Lines: 317-325). 

Comment:  Line 346: For Figure 4, there was a significant decline in water area in in-situ data whereas Sentinel-1 SAR didn’t capture the additional water drawdown? What is a potential reason for that?

Response: Thank you for raising this excellent observation. There are a few potential reasons that could explain why Sentinel-1 did not detect further declines in water area as observed in the in-situ water level data logger.

1. Slope of the exposed shoreline: Steeper sloped areas may not translate into proportional surface area reductions, so the satellite is unable to detect small water level drops. Gradual slopes are required to detect minor water level changes through area changes.
2. Shoreline bathymetry: Similarly, if the nearshore lake bottom profile is steep, moderate water level declines may not expose much additional area.
3. Frozen conditions: If drawdown continues into months when ice cover starts to form, the frozen lake surface may mask the open water-ice boundary that Sentinel-1 uses to estimate area. This can prevent detection of additional area loss as water levels decline under ice.

While we tried to account for these factors in our analysis, there are inherent challenges translating water level changes to area changes through remote sensing. In the revised manuscript, we have expanded the discussion of these factors and uncertainties to help explain potential discrepancies compared to in-situ data (Section 4.1; Lines: 567-576). We appreciate you highlighting this limitation.

Comment:  Line 355-358: Could you please elaborate on the physical reasons for why Sentinel-1 is showing the water area returned to normal before the in-situ water levels. Is this consistent across different years and across different lakes?

Response: We appreciate the reviewer highlighting this point about the time lag we observed between the satellite-derived surface water area recovery and the in-situ water level recovery in spring. We checked other lakes and different years and time lag varies from lake to lake based on lake size, bathymetric profile and shoreline slope of the lakes. The primary physical reason for such variable time lag could be the in-situ sensor deployment location because in-situ sensors provide measurements at a discrete point location, while the satellite image provides water area variation for overall lake. If water levels are increasing first in shallow near shore regions distant from the sensor, this could introduce a lag before the in-situ gauge starts to record the rising water. We found that for most individual lakes there was a time lag across years. However, the magnitude of lag time differed among lakes. Overall, these observations indicate that both lake-specific morphometry and spatial variable water level changes likely contribute to the detected time lags between satellite area recovery and in-situ water level recession each spring following drawdown. We have expanded the discussion of this phenomenon in the revised version of the manuscript to provide further insight into the factors driving the discrepancy (Section 4.1; Lines: 594-601).

Comment:  Line 371-372: This comment again brings the question of not deseasonalizing and detrending the time series. Have the authors tried to compare the results of 12-months and 6-months WD and ND on deseasonalized and detrended time series and compared it with 3-months results? Since in 3-month there is no inherent seasonality but for 12-month there is going to be strong interannual seasonality and a weaker but still present seasonality for 6-month time series.

Response: We appreciate the reviewer raising this insightful point. Testing the classification accuracy using deseasonalized and detrended time series versus raw data would provide a useful analysis to evaluate our methodology. Within the study’s timeframe, we are unfortunately unable to perform these additional experiments. However, we agree it would be valuable future work to explicitly compare raw versus processed time series as input to the classification algorithm. Our rationale for using the raw time series data was that the inherent seasonal decline during fall-winter months provide a key signal differentiating winter drawdown versus non-drawdown lakes. Removing these features through data processing may degrade the classifier performance, But we acknowledge this hypothesis requires testing to quantify the impact on accuracy. For the present study, we opted to retain the original seasonal and trend characteristics of the time series to best leverage the distinct signatures of drawdown hydrologic regimes. However, comparative testing of raw versus detrend/deseasonalized data would strengthen the methodology. We have noted this analysis as a priority for follow-up research to supplement the findings from our current approach. We appreciate you helping us think through options to thoroughly respond to this point.

Comment:  Line 526-528: Since machine learning algorithms are sensitive to unequal sample sizes, the accuracy results provided in the manuscript would be biased in favor of larger sample size. I understand the limitation of availability of in-situ data, but I would suggest the authors to provide a discussion on how it would affect the results in identifying WD vs ND lakes using kNN. You can add a few more sentences in section 4.2 at the end of the first paragraph discussing this limitation in context of your present results.

Response: We appreciate the reviewer raising this important point about the potential for biased results due to unequal class sample sizes in our kNN classification approach. We agree that this is an important limitation to acknowledge and needs to be added in the discussion. The reviewer is correct that with machine learning techniques like kNN, classification performance can be disproportionately influenced by the larger class when sample sizes between classes are imbalanced. In our case, the number of winter drawdown samples was greater than the non-drawdown samples in both the in-situ and satellite-based training data. This imbalance could result in biased results as the model is more heavily optimized to the dominant drawdown class. The issue of imbalance data is a common challenge with classification of real-world datasets. However, we agree the disproportionate sample sizes between classes should be highlighted as a potential source of bias. Here is a our addition to the Discussion section 4.2 (Lines: 620-628) to acknowledge this limitation:

“A limitation of the machine learning model development was the imbalance between the number of WD and ND samples available for training and testing the classification. The disproportionate sample sizes between classes may have biased the accuracy results in favor of the majority WD class. While we used cross-validation approaches to account for the imbalance, it may introduce some bias favoring the majority WD class. Increasing the ND sample size in future could improve accuracy, particularly the false positive rate. However, the overall accuracy achieved even with imbalance classes demonstrates the potential of machine learning for automated WD classification using remote sensing time series data.”

Round 2

Reviewer 2 Report

The authors have addressed all my concerns and I believe it is ready for publication.