Comparison of multi-source satellite remote sensing observations for monitoring the variations of small lakes: a case study of Dai Lai Lake (Vietnam)

The world has about 117 million lakes and reservoirs (Verpoorter et al. 2014), and despite covering only a tiny portion of the Earth's surface, they have a substantial influence on global hydrological equilibrium, as evidenced by their role in regulating global water balance (Ji et al. 2018). These water bodies also play an important role on the emission of atmospheric carbon dioxide and methane (Williamson et al. 2009). In addition to their environmental contributions, lakes and reservoirs offer significant benefits for human survival, ecosystem services, and economic prosperity (Pham-Duc et al. 2022; Vári et al. 2022). Over the last several decades, under the combined effect of climate change and human water withdrawals for agriculture, industry and domestic consumption, notable transformations in the expansion and declining of global lakes and reservoirs have been witnessed globally (Smith et al. 2005; Wurtsbaugh et al. 2017; Yao et al. 2023). This dynamic situation requires frequent and precise monitoring of spatial and temporal variations of lakes and reservoirs to provide evidence for better water management and environmental monitoring, particularly in developing countries (Steveson et al. 2010; Du et al. 2011; Zeng et al. 2017; Asfaw et al. 2020).

Several lake databases exist that provide insights of the dynamics of lakes and reservoirs from regional to global scales, such as the Global Surface Water dataset (Pekel et al. 2016), the Global Lakes and Wetlands database (Lehner & Döll 2004), the global abundance and size distribution of water bodies (Downing et al. 2006), and the abundance and size distribution of lakes and reservoirs in the United States (McDonald et al. 2012). While global databases normally are not effective in mapping lakes with a size smaller than 10 km², regional databases, on the other hand, offer details of small water bodies but are limited to much smaller areas (Hanson et al. 2007). Satellite remote sensing observations offer a unique and effective approach for mapping and monitoring water body dynamics across varied environments from the tropics to the poles, often employing spectral water indices such as the Normalized Difference Water Index (NDWI) (Gao 1996; McFeeters 1996), the Modified of Normalized Difference Water Index (MNDWI) (Xu 2006), or the Automated Water Extraction Index (AWEI) (Feyisa et al. 2014).

Since the launch of Landsat-1 satellite in 1972, Landsat series are the most widely used optical remote sensing dataset for lake mapping and monitoring (Pekel et al. 2016; Asfaw et al. 2020). However, its 16-day revisit time limits real-time and continuous mapping of water body dynamics. Since 2000, the launch of the Moderate Resolution Imaging Spectroradiometer (MODIS) has enabled daily observations of the Earth's surface, but its coarse spatial resolution of 500 m makes it not suitable for mapping small lakes and reservoirs (Feng et al. 2012). Sentinel-2 satellites provide higher spatial resolution observations (10 m) as well as temporal resolution (maximum 5 days), but their data is limited from 2014 when the first Sentinel-2A satellite was launched. Recently, commercial satellite observations acquired by PlanetScope (PS) satellites have shown their great potential for lake monitoring due to its high spatial resolution of 3 m and daily temporal coverage (Qayyum et al. 2020; Perin et al. 2021). However, cloud cover remains a major obstacle for the use of optical remote sensing, which limits their applications in many regions like the tropics (Pham Duc & Tong Si 2021).

To overcome this issue, using observations acquired by Synthetic Aperture Radar (SAR) satellites can be a solution because radar sensors are not affected by clouds and weather conditions. Since the launch in 2014, SAR Sentinel-1 observations have been used extensively for surface water monitoring and lake monitoring (Huth et al. 2020; Pham-Duc & Nguyen 2022). However, a comprehensive global radar-based surface water change dataset remains unavailable, and the first global permanent open water map, derived from 8 years of ENVISAT ASAR observations at 150 m spatial resolution, was published recently (Santoro et al. 2015).

Each satellite sensor has its own advantages and limitations; therefore, their capability for lake mapping and monitoring is not similar. The main objective of this study is to compare the efficacy of different satellites for monitoring small lakes as there are only a few articles provided these comparisons (Mullen et al. 2023). Radar and free-cloud satellite imagery from three satellites, namely Sentinel-1, Sentinel-2, and PS, were utilized for the comparison. Due to the absence of a ground truth dataset for comparison with results obtained from satellite observations, it would be challenging to conclude which satellite product and method provide the most accurate estimation, but it is worth investigating the potential of each dataset. Dai Lai Lake, located in the North Vietnam, was selected as the study area, and the time span is limited to the 2018–2023 period when observations from all three satellites were available and accessible.

This paper is structured as follows: Section 2 describes details of the study area and datasets used. The methodology is presented in Section 3, and results are shown in Section 4 and discussed in Section 5. Section 6 concludes this study.

Sentinel-1 is a European Union (EU)-funded satellite project, falls under the Copernicus program and is implemented by the European Space Agency (ESA). The mission comprises two identical SAR satellites: Sentinel-1A and -1B, launched in April 2014 and 2016, respectively. Regrettably, on December 23, 2021, Sentinel-1B encountered an anomaly related to the instrument electronics power supply unit, crucial for providing power to the radar electronics (ESA 2022). Consequently, ESA and the European Commission announced the termination of the Sentinel-1B mission, and plans are in force to expedite the launch of the Sentinel-1C at the earliest opportunity. Sentinel-1 satellites operate at C-band (5.405 GHz), with an incidence angle ranging from 29° to 46°. Positioned in near-polar sun-synchronous orbits at an altitude of approximately 693 km, each Sentinel-1 satellite offers a temporal resolution of 12 days. Sentinel-1's capabilities enable a wide range of monitoring applications for both land and ocean surfaces in all-weather conditions, including day and night (ESA 2015). In this study, the author has only utilized Sentinel-1A Ground Range Detected (GRD) Level-1 observations from the Interferometric Wide (IW) swath mode with a spatial resolution of 5 m × 20 m. To save time for data downloading and processing, this study utilized Sentinel-1 observations stored in the Google Earth Engine (GEE) platform, where they were pre-processed before being uploaded to the cloud. The pre-processing of Sentinel-1 imagery in GEE follows five main basic steps (ESA 2016). Firstly, the orbit metadata is updated and corrected for each acquisition. Secondly, GRD border noise removal is applied to eliminate invalid signals and noise along the image edges. Thirdly, thermal noise removal is applied to reduce disconnections between sub-swaths within the images. Next, radiometric calibration is performed to compute the backscatter coefficient. Finally, terrain correction is applied to compensate for the distortions caused by the side-looking geometry of the satellites.

ESA's Sentinel-2 mission is another satellite project belonging to the Copernicus program, which comprises two sun-synchronous satellites named Sentinel-2A (launched in June 2015) and Sentinel-2B (launched in March 2017). Orbiting at an average altitude of 786 km, these two optical satellites provide a 5-day revisiting time at the equator. The key instrument onboard Sentinel-2 satellites is the Multispectral Instrument (MSI) sensor, which encompasses 13 spectral bands ranging from visible (400 nm) to short-wave infrared wavelengths (SWIR; 2,200 nm). Depending on the wavelengths, these two satellites offer three distinct spatial resolutions of 10, 20, and 60 m. In this study, the author utilized Sentinel-2 Level-2A Bottom-Of-Atmosphere (BOA) reflectance imagery, also stored in the GEE platform. To detect and monitor the variations of the lake's surface water area, both the NDWI and the MNDWI have been utilized, using data from band 3 (green), band 8 (NIR) and band 11 (SWIR), respectively.

The PS constellation, operated by Planet Labs Inc., consists of small commercial satellites capable of daily imaging the entire Earth's land surface. The PS constellation encompasses three generations of satellites: Dove Classic (available from July 2014 to April 2022), Dove-R (available from March 2019 to April 2022), and SuperDoves (available from March 2020 to the present time) (PlanetLabs 2023b). Each generation comprises multiple ‘flocks’, with each flock consisting of satellites launched and placed into similar orbits around the same time. All PS satellites orbit at approximately 475 km in a sun-synchronous orbit with an inclination angle of 98⁰. They provide high spatial resolution imagery in four different bands: red, green, blue, and NIR. In this study, the author utilized free-cloud PS Level 3B Ortho Analytic surface reflectance imagery at 3 m spatial resolution to monitor the variations in the surface water area of the lake, using data from band 2 (green) and band 4 (NIR). Despite being a private company, Planet Labs offers its PS observations for research and education purposes through the Planet's Education and Research Program (PlanetLabs 2023a). For detailed technical information about the PS constellation and their data, readers can refer to previous papers (Frazier & Hemingway 2021).

The comparative spatial comparison of surface water extent maps, derived from the three satellite sensors, is presented in Figures 4 and 5, revealing both the minimal and maximal disparities between these maps. In Figure 4, all satellite acquisitions took place in December 2021, showcasing a similar spatial distribution of the lake across observations. The estimated surface water area, utilizing NDWI and MNDWI values from the Sentinel-2 satellite, stands at 371.76 and 373.11 ha, respectively, while the estimation derived from the Sentinel-1 image is 361.06 ha. The surface water area estimated using PS image reaches 375.87 ha, representing a marginal 0.74 and 1.1% increase over results from Sentinel-2 satellite, and a 4.1% increment compared to results from Sentinel-1 satellite. Figure 5 illustrates the spatial distribution of the lake, acquired on November 5, 2019. Similar to Figure 4, the lake's outline remains comparable, through the differences are more pronounced. The total estimated surface water area derived from Sentinel-2 NDWI and MNDWI reaches 369.42 and 370.36 ha, respectively, while the Sentinel-1 image results an estimate of 351.14 ha. Results derived from PS image return a larger surface water area of 382.62 ha, marking a 4.5 and 10.1% increase over Sentinel-2 and Sentinel-1 outcome, respectively. The discrepancies of the result water maps primarily occur at two regions: firstly, the land–water border of the lake where the higher spatial resolution of PS observations provides enhanced land–water border delineation, and secondly, near the northern lake branch where shallow water causes misclassification to our methodology.

The proposed method has demonstrated its efficacy in mapping and monitoring small lakes through the utilization of both optical and radar satellite observations. However, it is crucial for scholars to be mindful of specific phenomena which might occur when embarking on future research in this direction. In the following subsection, the author focuses on two noteworthy phenomena that can significantly impact accuracy. The first phenomenon is related to the resemblance between SAR Sentinel-1 backscatter coefficients over the lake surface and those over a nearly flat soil surface. The second one pertains to the presence of negative signals over water bodies in some PS-derived NDWI maps, which the author did not expect.

This work investigates the comparative capabilities of various satellite sensors, from radar Sentinel-1 to optical Sentinel-2 and PS imagery, for mapping and monitoring Dai Lai Lake, a small lake situated in North Vietnam, over the period from 2018 to 2023. Sentinel-1 and cloud-free Sentinel-2 observations were partly processed in GEE to derive the VH-polarized backscatter coefficient, and the NDWI and MNDWI maps, while cloud-free PS observations were entirely processed using local computers to produce the NDWI maps. The process of generating binary water/non-water maps involves mainly the application of the Otsu threshold selection algorithm to each individual map generated from the previous step. Results showed that Dai Lai Lake's surface water extent does not strongly follow the seasonal cycle; rather, it remains relatively stable throughout the 6-year period. The fluctuations between the minimal and maximal extents, ranging from 19 to 28 ha, depend on the sensor being utilized. Despite differences in sensor type, the spatial distribution outcomes of the lake exhibit great similarity. The lake's average surface water extent estimated from 39-m resolution PS imagery is about 2.17 and 5.60% more than that obtained from 10-m resolution of Sentinel-2 and Sentinel-1 imagery, respectively. PS observations are effective in small lake mapping and monitoring due to their high spatial resolution and daily revisiting frequency. However, it is advised that researchers pay attention to the quality of its NIR wavelength. Meanwhile, Sentinel-2 observations demonstrate great potential for lake monitoring, with similar results obtained using both the NDWI and MNDWI. On the other hand, the strength of Sentinel-1 satellite lies in its all-weather operational capability. Yet, the reliance on a single VH band could result in misclassification due to similarities in backscatter coefficients between water surfaces and other flat surfaces.

One significant limitation of this study is the absence of a ground truth dataset for result comparison against satellite observations. The absence of a reference dataset makes it challenging to determine which dataset and method perform best. Each satellite sensor has distinctive advantages and constraints. Therefore, to effectively monitor spatial–temporal changes in lakes over a long period, it is recommended to harness the combined capabilities of all available satellite observations. The choice of the primary satellite dataset and methodology to be employed greatly depends on the researchers and the study areas. In regions with low cloud contamination, such as Africa, optical satellite observations should be preferred. In contrast, for areas where cloud contamination is high, particularly in tropical regions, radar satellite observations are a more suitable choice. When dealing with small regions, high spatial resolution imagery from PS satellites is more appropriate, while for larger areas, observations provided by other satellite platforms with wider swath coverage should be the primary datasets.

The author would like to thank ESA for providing Sentinel-1 and Sentinel-2 observations. PS observations were provided by Planet Labs through the Planet's Education and Research Program. The author also thanks Google for providing the GEE platform, which was utilized for pre-processing Sentinel-1 and Sentinel-2 observations. The author would like to thank the editor and four reviewers for their valuable comments and suggestions, which greatly improved the quality of this manuscript.

This research was funded by the Vietnam Academy of Science and Technology (VAST), grant number THTEXS.03/22-24 to Binh Pham-Duc.

B. P.-D. contributed to the study conception and design, material preparation, data collection and analysis, and wrote the manuscript.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.