Enhanced Remote Sensing of Inland Water Surface Elevation Using Sentinel-3 Radar Altimeter (SRAL)

Enhanced Remote Sensing of Inland Water Surface Elevation Using Sentinel-3 Radar Altimeter (SRAL) Mahdis Rezapour* Department of Photogrammetry and Remote Sensing, Faculty of Geomatics Engineering, K. N. Toosi University of Technology, Iran. [email protected] * Contributed equally to this paper. Elahe Khesali Department of Photogrammetry and Remote Sensing, Faculty of Geomatics Engineering, K. N. Toosi University of Technology, Iran. [email protected] Amir Naghibi Division of Water Resources Engineering, Faculty of Engineering (LTH), and Centre for Advanced Middle Eastern Studies (CMES), Lund University, Sweden. [email protected] Mohammad Javad Valadan Zoej Department of Photogrammetry and Remote Sensing, Faculty of Geomatics Engineering, K. N. Toosi University of Technology, Iran. [email protected] Ali Mehran Department of Civil and Environmental Engineering, San Jose State University, California, USA. [email protected] Hossein Hashemi Division of Water Resources Engineering, Faculty of Engineering (LTH), and Centre for Advanced Middle Eastern Studies (CMES), Lund University, Sweden. [email protected] Alireza Taheri Dehkordi* Division of Water Resources Engineering, Faculty of Engineering (LTH), Lund University, Sweden. [email protected] * Contributed equally to this paper. Alireza Farahmand Department of Geography, Geology, and Environment, California State University, California, USA [email protected] Abstract— Inland surface waters are vital for the global carbon, energy, and water cycles; however, they are affected by climate change and human activities. Continuous monitoring of Water Surface Elevation (WSE) is essential for assessing water quantity and ensuring sustainable management. Traditional insitu WSE measurements often face high costs and accessibility challenges. This study examines the use of satellite altimetry, specifically from the Sentinel-3 (S3) Radar Altimeter (SRAL), to enhance WSE estimates through Machine Learning (ML) techniques, the Random Forest (RF) algorithm. By applying the RF algorithm to SRAL data over Lake Michigan, we corrected altimeter-derived WSE values against in-situ data. Our results show a reduction in the Root Mean Squared Error (RMSE)— from 35 cm for the mean WSE across all Virtual Station (VS) points to 9 cm for the RF-corrected WSE—when compared against in-situ measurements using a Leave-One-Date-Out (LODO) approach. Additionally, the R-squared (R2) improved from 0.88 to 0.96, indicating a stronger correlation between the corrected WSE and in-situ measurements. These findings highlight the effectiveness of the ML method in improving the accuracy of altimetry-based WSE estimation. Moreover, this research underscores the potential of integrating ML methods with remote sensing techniques for enhanced inland water resource management. Keywords—Remote Sensing, Satellite Altimetry, Machine Learning, Random Forest, Sentinel-3. I. INTRODUCTION Inland surface waters, significantly impacted by climate change and human activities, play a crucial role in the global carbon, energy, and water cycles [1]. They also provide essential resources for the domestic, agricultural, and industrial sectors. Consequently, continuous monitoring of these resources is of paramount importance to understand their dynamics and ensure sustainable management. Among various hydrological parameters, Water Surface Elevation (WSE) is particularly important for assessing the quantity of inland surface waters. Traditionally, WSE has been measured through in-situ gauging stations. However, high installation and maintenance costs, combined with the inaccessibility of remote or steep terrains, highlight the need for alternative approaches [2]. Satellite altimetry, is an active Remote Sensing (RS) technique that measures the round-trip travel time of an electromagnetic pulse between a satellite and the Earth’s surface. This method provides a promising solution for global coverage and repeated observations [3]. Although altimeters were initially designed for oceanographic applications, advances in satellite altimetry have now made continuous WSE monitoring of inland waters feasible. This is achieved through repeat-track orbits that revisit the same locations at regular intervals [4]. Additionally, two developments have further improved inland WSE monitoring: 1) Open-Loop Tracking Command (OLTC), first implemented in the Jason2 mission, which positions the tracking window using precomputed altitudes stored onboard in a Digital Elevation Model (DEM) [5]; and 2) Synthetic Aperture Radar (SAR) mode with Delay-Doppler processing (pioneered by CryoSat2). This mode reduces random noise and increases along-track resolution compared to the Low Resolution Mode (LRM) [6]. Sentinel-3 (S3), part of the European Space Agency (ESA) Copernicus program, provides multi-sensor observations of Earth’s surface, including an altimeter, called Sentinel-3 Radar Altimeter (SRAL) that operates in both OLTC and SAR modes [7]. This makes S3 particularly suitable for inland WSE retrieval. However, even with these advancements, accurate inland WSE estimation via satellite altimetry remains 8328979-8-3315-0810-4/25/$31.00 ©2025 IEEE IGARSS 2025 IGARSS 2025 - 2025 IEEE International Geoscience and Remote Sensing Symposium | 979-8-3315-0810-4/25/$31.00 ©2025 IEEE | DOI: 10.1109/IGARSS55030.2025.11243573 Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 05,2025 at 17:57:41 UTC from IEEE Xplore. Restrictions apply. challenging. Discrepancies often arise between altimeterderived WSE and in-situ measurements due to atmospheric effects, sensor-based uncertainties, retracking errors, and waveform contamination over inland water bodies where large footprints can capture signals from land or vegetation [8]. These factors directly affect the accuracy of WSE estimates, emphasizing the need to improve satellite altimetry performance for inland WSE monitoring. Numerous studies have evaluated the performance of different satellite altimeters over inland water bodies [9, 10]. However, to the best of our knowledge, this paper is the first to explore enhancing altimeter-derived WSE through Machine Learning (ML) approaches. ML methods are particularly valuable for modeling non-linear and complex relationships in data, among which Random Forest (RF) has demonstrated strong performance across various applications [11]. In this study, we employ SRAL data to improve inland WSE estimates in Lake Michigan (LM) in the US, by using RF to correct the altimeter-derived WSE. The remainder of this paper is organized as follows: Section II introduces the study area and datasets used, Section III outlines the proposed methodology, Section IV presents and discusses the results, and Section V concludes the study. II. STUDY AREA AND DATASETS A. Study Area Lake Michigan (LM) was selected as the study area for this research. As one of the Great Lakes in North America (Fig. 1a), LM is a vital freshwater resource with ecological, economic, and social importance. LM ranks as the sixthlargest freshwater lake in the world, making it a critical focus for hydrological studies. Recent trends, such as water-level fluctuations driven by changes in precipitation patterns and evaporation rates, underscore the need for continuous monitoring of WSE and adaptive management strategies [12]. B. Datasets S3 is a constellation comprising two identical satellites, S3A and S3B, which fly in an orbit nearly identical to Envisat, thereby allowing the continuation of the ERS/Envisat time series [7]. S3A was launched in February 2016, followed by S3B in April 2018, both placed in a sun-synchronous orbit at an altitude of approximately 814 km with an inclination of 98.65°. The mission includes a dual-frequency altimeter with a single nadir-looking antenna. Ku-band serves as the primary frequency for range measurements, while C-band is used for ionospheric corrections. S3 has an along-track resolution of about 300 meters and a 27-day repeat cycle. In this study, we utilized Land (LA) Non-Time Critical (NTC) Hydrological (HY) Thematic products acquired by S3B in SAR mode at Kuband with a 20 Hz sampling rate. These data, were downloaded from the Copernicus website (https://dataspace.copernicus.eu, last accessed on 1 January 2025) for pass number 663 over LM, which represents the closest S3 ground track across this lake (Fig. 1b). We used enhanced Level-2 data files, which include geophysical corrections. The data were acquired on 73 different dates, spanning from late 2018 to the end of 2024. To validate SRAL performance and to train and evaluate the RF model for enhanced WSE estimation, we used in-situ measurements from the nearest United States Geological Survey (USGS) gauging station (Fig. 1b) along the S3B ground track. This station provided long-term WSE observations at a 1-second sampling rate (https://maps.waterdata.usgs.gov, last accessed on 1 January 2025). As a result, altimeter-based and in-situ WSE measurements were matched in time due to the station’s high sampling frequency. III. PROPOSED METHODOLOGY In satellite altimetry, WSE is determined by measuring the time it takes for an emitted pulse to travel from the satellite to the water surface and back [13]. The SRAL operates in the microwave range, allowing it to measure WSE both during the day and night without being affected by meteorological conditions. To accurately compute WSE, three essential components are required: 1) the satellite’s precise location to derive its altitude (H) relative to a reference datum, which is derived by advanced onboard navigation systems; 2) estimating the round-trip travel distance of the pulse to derive the range (R) between the satellite and the water surface; and 3) applying corrections to the altimetric measurements to ensure accuracy. Equation (1) provides the formula for computing the water surface elevation (WSE) using the aforementioned components, where WSESRAL is the SRALestimated WSE, H represents the satellite altitude, R denotes the distance between the satellite and the water surface (derived from the empirical Offset Center of Gravity (OCOG) Fig. 1. (a) Great Lakes in North America, (b) S3B track over LM and closest USGS in-situ gauging station (pink circle), providing 1-second measurements of WSE. 8329 Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 05,2025 at 17:57:41 UTC from IEEE Xplore. Restrictions apply. retracking algorithm in the standard S3 products), and Corr encompasses ionospheric, wet and dry tropospheric, solid Earth tide, and ocean tide corrections [14]. WSESRAL = H – (R + Corr) () All WSE values in this study were referenced to the WGS84 ellipsoid. As shown in Figure 1b, satellite altimetry provides WSE estimates at Virtual Station (VS) points, which are the intersection of the satellite’s ground track with a water body. Over LM, SRAL yielded 567 VS points, each associated with a WSESRAL measurement. These measurements served as input features to an RF model, where the in-situ gauge measurement for each date acted as the target variable. The RF model was selected for its robustness against noise and outliers, thanks to ensemble learning; its computational efficiency, as it does not require many hyperparameters; and its proven performance in various hydrological RS applications compared to other ML models [11]. The RF algorithm is an ensemble of decision trees [15]. Each tree is trained on a different subset of the data, and final predictions are obtained by averaging the individual tree outputs. To determine the optimal hyperparameters, we performed a grid search with 5-fold cross-validation on a random 70% subset of the entire dataset. There were 73 observation dates (corresponding to the SRAL acquisitions), yielding 73 data points—each consisting of 567 independent variables (WSESRAL measurements) and one dependent variable (the in-situ WSE (WSEin-situ) from the gauge station). The grid for hyperparameter tuning included: 1) n_estimators (number of trees) ranging from 50 to 300 in steps of 50, 2) min_samples_split (minimum samples required to split an internal node) from 2 to 5 in steps of 1, and 3) min_samples_leaf (minimum samples required at a leaf node) from 1 to 5 in steps of 1. After determining the optimal hyperparameters, we computed three Root Mean Squared Error (RMSE) and three R-squared (R2) values: 1) RMSESRAL and R2SRAL, between the mean WSESRAL (across all VS points) and WSEin-situ, 2) RMSERF and R2RF, between the RFenhanced WSE (WSERF) and WSEin-situ, and 3) RMSEavgcorr and R2avgcorr derived by applying an average bias correction — based on discrepancies between the mean WSESRAL across all VS points and the corresponding WSEin-situ —to the mean WSESRAL at each date. RMSERF and R2RF were computed using a Leave-One-Date-Out (LODO) evaluation method to assess the RF model's ability to generate the time series of WSE. IV. RESULTS AND DISCUSSION As discussed in the previous section, a 5-fold cross-validation on a random 70% subset of the entire dataset was employed for hyperparameter tuning of the RF model. Based on these experiments, the optimal RF hyperparameters were: ('n_estimators': 200, 'min_samples_split': 2, 'min_samples_leaf': 3) [11]. Fig. 2. compares the values of insitu WSE (WSEin-situ) against the three scenarios introduced in Section III. As shown, the RMSESRAL is 35 cm, while applying an average bias correction reduces the RMSEavgcorr to 18 cm. However, using the RF model significantly improves the WSE estimates, yielding an RMSERF of 9 cm. The corresponding R² values for these methods are R²SRAL= 0.88, R²avgcorr = 0.90, and R²RF = 0.96. Moreover, as shown in the third scenario (RFbased improvement), the density of points around the perfect fit line (y = x) is the highest compared to the other two scenarios. These results clearly demonstrate that correcting WSESRAL using ML approaches such as RF can substantially enhance measurement accuracy. Fig. 2. Comparison of WSEin-situ values against different scenarios using LODO approach. Point densities describe how densely populated the space is around each point. Higher values indicate that a point is in a region with many nearby neighbors, while lower values suggest it lies in a more isolated or sparse area. To make the point densities comparable across all three plots, the values were normalized to a range between 0 and 1. Fig. 3 displays the time series of WSE values derived from different approaches for 73 dates spanning 2018 to 2024. The boxplots represent the distribution of all WSESRAL measurements across VS for each date. As evident, the raw SRAL distribution often does not perfectly align with in-situ measurements. While average bias correction provides a moderate improvement, the RF-enhanced WSE values exhibit substantial consistency with the WSEin-situ, as also illustrated in Fig. 2. The results revealed that the RF model outperformed the simple average bias correction method. This illustrates the advantages of using ML to capture the non-linear and complex relationships between altimeter-derived WSE and in-situ observations. While the average bias correction moderately improved WSE estimates, its reliance on simple linear adjustments highlights its limitations in capturing the full complexity of the relationships. In contrast, the RF model demonstrated its ability to deliver consistently accurate results across multiple dates. Having said that, this study focused on LM and the proposed approach could be extended to other regions. However, the availability of high-quality in-situ gauge data remains an important consideration for achieving comparable results. Future work could explore the application of similar methodologies to other lakes and water bodies globally, as well as incorporating additional satellite missions, such as Sentinel-6, to further improve monitoring accuracy and temporal coverage. Testing advanced ML models, such as Deep Learning (DL) architectures, could further enhance performance, especially for complex hydrological scenarios [16]. Furthermore, applying this methodology to rivers, smaller lakes, or reservoirs with more dynamic topographies and exploring its robustness would provide valuable insights. 8330 Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 05,2025 at 17:57:41 UTC from IEEE Xplore. Restrictions apply. V. CONCLUSION In this study, we demonstrated the effectiveness of integrating satellite altimetry with ML techniques to enhance inland WSE measurements. Using S3 SRAL data and RF model, the accuracy of WSE estimations were improved, compared to in-situ gauge data. The results showed a notable reduction in RMSE for raw SRAL-derived WSE compared to RF-enhanced WSE, along with an improvement in the Rsquared (R2), indicating the potential of ML models to address discrepancies caused by atmospheric effects, retracking errors, and waveform contamination in satellite altimetry. Additionally, the RF model consistently outperformed simple average bias correction method, which highlights the advantage of leveraging ML's ability to model non-linear relationships between altimetry-derived WSE and in-situ measurements. The proposed methodology of this paper is a scalable and adaptable framework for improving satellite altimetry data, making it suitable for use in other inland water bodies with similar data availability. REFERENCES [1] Z. Chen et al., "Remote sensing research on plastics in marine and inland water: Development, opportunities and challenge," Journal of Environmental Management, vol. 373, p. 123815, 2025. [2] S. C. Palmer, T. Kutser, and P. D. Hunter, "Remote sensing of inland waters: Challenges, progress and future directions," vol. 157, ed: Elsevier, 2015, pp. 1-8. [3] S. Abdalla et al., "Altimetry for the future: Building on 25 years of progress," Advances in Space Research, vol. 68, no. 2, pp. 319-363, 2021. [4] S. Kossieris, V. Tsiakos, G. Tsimiklis, and A. Amditis, "Inland Water Level Monitoring from Satellite Observations: A Scoping Review of Current Advances and Future Opportunities," Remote Sensing, vol. 16, no. 7, p. 1181, 2024. [5] J. Lambin et al., "The OSTM/jason-2 mission," Marine Geodesy, vol. 33, no. S1, pp. 4-25, 2010. [6] L. Jiang, R. Schneider, O. B. Andersen, and P. Bauer-Gottwein, "CryoSat-2 altimetry applications over rivers and lakes," Water, vol. 9, no. 3, p. 211, 2017. [7] C. Donlon et al., "The sentinel-3 mission: Overview and status," in 2012 IEEE International Geoscience and Remote Sensing Symposium, 2012: IEEE, pp. 1711-1714. [8] Z. Huang, R. Sun, H. Wang, and X. Wu, "Trends and Innovations in Surface Water Monitoring via Satellite Altimetry: A 34-Year Bibliometric Review," Remote Sensing, vol. 16, no. 16, p. 2886, 2024. [9] A. Nalluri, H. Ramesh, and P. R. Dhote, "Monitoring Water Level Fluctuations of Reservoirs in The Krishna River Basin Using Sentinel3 and ICESat-2 Altimetry Data," IEEE Access, 2024. [10] B. Nilsson and K. Nielsen, "Validation of Sentinel-6MF based lake levels–An assessment with in situ data and other satellite altimetry data," Advances in Space Research, vol. 73, no. 12, pp. 5806-5821, 2024. [11] A. Taheri Dehkordi, M. J. Valadan Zoej, H. Ghasemi, E. Ghaderpour, and Q. K. Hassan, "A new clustering method to generate training samples for supervised monitoring of long-term water surface dynamics using Landsat data through Google Earth Engine," Sustainability, vol. 14, no. 13, p. 8046, 2022. [12] V. Y. Cheng, A. Saber, C. A. Arnillas, A. Javed, A. Richards, and G. B. Arhonditsis, "Effects of hydrological forcing on short-and long-term water level fluctuations in Lake Huron-Michigan: A continuous wavelet analysis," Journal of Hydrology, vol. 603, p. 127164, 2021. [13] L. Yang et al., "Satellite altimetry: Achievements and future trends by a scientometrics analysis," Remote Sensing, vol. 14, no. 14, p. 3332, 2022. [14] L. Jiang, K. Nielsen, S. Dinardo, O. B. Andersen, and P. BauerGottwein, "Evaluation of Sentinel-3 SRAL SAR altimetry over Chinese rivers," Remote Sensing of Environment, vol. 237, p. 111546, 2020. [15] L. Breiman, "Random forests," Machine learning, vol. 45, pp. 5-32, 2001. [16] A. Taheri Dehkordi, H. Hashemi, A. Naghibi and A. Mehran, "Ensemble of Pruned Bagged Mixture Density Networks for Improved Water Quality Retrieval Using Sentinel-2 and Landsat-8 Remote Sensing Data," IEEE Geoscience and Remote Sensing Letters, vol. 21, pp. 1-5, 2024. Fig. 3. Time series of WSE estimates for 73 data acquisitions (2018–2024). Boxplots show the distribution of all WSESRAL measurements across all VS points. Black circles indicate in-situ measurements, red circles the RF-enhanced WSE (WSERF), and green circles the WSE after average-correction approach (WSEavgcorr). 8331 Authorized licensed use limited to: Universita degli Studi di Roma Tor Vergata. Downloaded on December 05,2025 at 17:57:41 UTC from IEEE Xplore. Restrictions apply.