Satellite Imagery-Based Monitoring of Water Surface Variability and Its Relationship with Rainfall at Baengnokdam, Mount Hallasan, Jeju Island, South Korea

Abstract

This study quantitatively analyzed the temporal changes in the surface area of the Baengnokdam Crater Lake, located at the summit of Mount Hallasan on Jeju Island, using Planet Scope satellite imagery, and evaluated its correlation with precipitation and water level. A total of 227 satellite images acquired between April and November from 2017 to 2024 were used to calculate the Normalized Difference Water Index (NDWI) and delineate the waterbody boundary of Baengnokdam. The NDWI-based analysis revealed distinct seasonal variability, with peak water surface areas frequently observed in May and September. To examine the temporal lag between rainfall events and the surface area response, the cross-correlation function was applied. The highest correlation (r = 0.496)was observed at a 4-day lag, suggesting the possibility of a time-lagged hydrological response. Regression analysis between water level and satellite imagery-derived surface area using data from 2021, 2022, and 2024 showed that the linear regression model yielded an average coefficient of determination of R² = 0.741, outperforming the logarithmic model. The results demonstrated that an increase of 1,000 m² in surface area corresponds to an average water level rise of approximately 24.1 cm. These findings demonstrate the applicability of time-series satellite imagery for exploring alpine lakes' responses to rainfall and for investigating the relationship between surface area and water level.

1. Introduction

Mount Hallasan, located at the center of Jeju Island—a volcanic island situated at the southern end of the Korean Peninsula—hosts a crater at its summit formed by volcanic eruptions. Within this crater, a low-permeability layer composed of volcanic ejecta and sediment has resulted in the formation of a lake, known as Baengnokdam Crater Lake (Ko, 2010). The impounded water of Baengnokdam does not persist year-round and fluctuates frequently depending on precipitation. During periods of heavy rainfall, the water body can expand to approximately 20,912 m², with an average depth of about 1.62 meters (Ko, 2006, 2010; Yim et al., 2018).

Baengnokdam holds not only geomorphological significance but also scenic, ecological, and educational value. It has been designated as Scenic Site No. 90 (Lee, 2014) and is part of the UNESCO World Natural Heritage. Due to its appeal as a tourist destination, many visitors hike to the summit of Hallasan, and the water body at Baengnokdam undergoes visually dramatic seasonal changes, leaving a strong impression. According to observation records from the Mount Hallasan Research Institute, heavy rainfall events caused by the northward migration of the monsoon front during summer or the passage of typhoons over Jeju Island raise the water level, creating a striking spectacle (Hallasan Ecological and Cultural Research Institute, Jeju Special Self-Governing Province, 2021b).

Numerous valleys extend from Baengnokdam at the summit to the coastline, contributing to the regional hydrology (Hallasan Ecological and Cultural Research Institute, Jeju Special Self-Governing Province, 2021a). Particularly, the waterbody of Baengnokdam provides essential habitat conditions for alpine flora and fauna, playing a key role in conserving endemic species. A decline in water levels, therefore, may directly lead to biodiversity loss in the alpine ecosystem. Previous studies documented a noticeable decline in water level, with the lakebed occasionally exposed during dry seasons (Research Institute for Hallasan, Jeju Special Self-Governing Province, 2013). This depletion phenomenon reportedly occurs for about 40 days annually (Ko, 2006).

Given these concerns, systematic monitoring and management of changes in the impounded water of Baengnokdam are crucial for preserving the landscape of Hallasan, maintaining Jeju’s ecosystems, conserving biodiversity, and protecting this valuable tourist resource. However, previous studies have focused on structural stability, such as preventing crater wall collapse or maintaining slope stability, and existing water level observations have been short-term (Research Institute for Hallasan, Jeju Special Self-Governing Province, 2013). Natural habitats across Jeju Island have declined markedly over the past few decades, leading to a general degradation in ecosystem quality; these changes have also affected the high-elevation ecosystems of Hallasan, raising concerns about the long-term sustainability of endemic species and habitats in the Baengnokdam area (Choi et al., 2021; Hong et al., 2021). Consequently, there is an increasing need for long-term conservation strategies and hydrological monitoring from an ecological perspective. Quantitative and long-term analyses remain insufficient. Some studies have suggested that sediment accumulation—together with rapid increases in permeability within the sediment layer and enhanced permeability along the lake margin caused by deposition of coarse, slope-derived grains—has contributed to the declining water level (Koh et al., 2003; Ko et al., 2009).

In difficult-to-access regions such as mountainous areas or crater lakes like Baengnokdam, satellite imagery-based waterbody analysis offers an effective alternative to in-situ observations. Mahto and Kushwaha (2018) conducted a time-series analysis of water surface area using Landsat imagery for the Lonar Crater Lake in Maharashtra, India, and quantified seasonal variations in surface area and water level. Chanda and Hossain (2024) applied the Normalized Difference Water Index (NDWI) to PlanetScope imagery, achieving over 90% accuracy in detecting waterbodies in narrow streams in Tennessee, USA, demonstrating its applicability for small waterbodies in mountainous regions. Chipman (2019) used multi-sensor satellite data, including Advanced Very High Resolution Radiometer (AVHRR), Moderate Resolution Imaging Spectroradiometer (MODIS), and Landsat, to monitor area, water level, and volume in Egypt’s Toshka Lakes over the long term, showing the effectiveness of such approaches for inaccessible areas.

In this study, PlanetScope imagery from multiple time points was used to extract water boundaries of Baengnokdam and analyze changes in the impounded water area. Subsequently, correlations between water surface area and rainfall were examined. Moreover, a relationship model between water level and area was established using field-based water level data, enabling estimation of water levels from satellite imagery in the absence of in-situ observations.

2. Materials

2.1. Study Area and Data Acquisition

2.1.1. Study Area

The study area selected is the Baengnokdam region of Mount Hallasan (Fig. 1). Baengnokdam is a crater lake located within a circular volcanic crater at the summit of Mount Hallasan (approximately 1,947 meters above sea level) on Jeju Special Self-Governing Province, situated at 33°21′31″N latitude and 126°32′10″E longitude (Ko, 1999). The Baengnokdam area is characterized by its high elevation, substantial variability in weather and climate conditions, and significant seasonal and precipitation-induced changes in water level and surface area. In particular, when the monsoon front stagnates around Jeju Island or when tropical cyclones (typhoons) pass near the region and move northward, the water level of Baengnokdam rises rapidly (Hallasan Ecological and Cultural Research Institute, Jeju Special Self-Governing Province, 2021b).

Fig. 1. Study area around Baengnokdam, Mount Hallasan, Jeju Island, South Korea. The area outlined in the red box in the right panel is a PlanetScope satellite image acquired on May 9, 2024. The red points in the image indicate the locations where water levels were measured. The red box indicates the region of interest defined for waterbody detection. The red star indicates the location of the Witseoreum AWS station (background image from Earthstar Geographics and Maxar).

2.1.2. Data

To detect the surface area of impounded water in Baengnokdam, PlanetScope imagery operated by Planet Labs (https://www.planet.com), which utilizes a constellation of Dove satellites, was employed. PlanetScope offers a spatial resolution of approximately 3–5 meters and provides a high revisit frequency, allowing daily or near-daily observations of the same location (Mullen et al., 2023). Through Planet Labs’ satellite imagery search system, Planet Explorer (https://www.planet.com/explorer), a total of 942 surface reflectance PlanetScope images with less than 40% cloud coverage were initially retrieved for the study period from January 2017 to December 2024. Among these, we selected images suitable for analysis based on cloud-free visibility of Baengnokdam and clear delineation of the land-water boundary. The selected images correspond to the 3B_AnalyticMS_SR product of PlanetScope, which is an atmospherically corrected, four-band satellite imagery product orthorectified for analytical applications (Planet Labs PBC, 2025).

During the winter season (December to March), the detection of impounded water was not possible due to snow cover and surface freezing. Therefore, only satellite imagery acquired between April and November was used, resulting in a total of 227 scenes for analysis alongside rainfall data (Table 1). Prior to data processing, the analysis area was spatially constrained to a 0.07 km² region that encompassed Baengnokdam in order to minimize potential misclassification of features with spectral characteristics similar to water bodies (Fig. 1).

Table 1. Number of PlanetScope satellite images used for analysis between April and November from 2017 to 2024

To compare and analyze the surface area of impounded water in Baengnokdam with rainfall data, Automatic Weather System (AWS) data were used. The AWS data were collected for the period from April to November through the Korea Meteorological Administration’s Weather Portal (https://www.weather.go.kr). Rainfall data for the Baengnokdam area were obtained from the AWS station at Witseoreum (33°21′44″N latitude and 126° 31′5″E longitude), which is located approximately 1.4 km from Baengnokdam.

2.2. Methodology

2.2.1. Calculation of the Water Surface Area of Baengnokdam

To identify and analyze changes in the impounded water surface area of Baengnokdam on Mount Hallasan, the NDWI, proposed by McFeeters (1996), was utilized. NDWI is calculated using the green and Near-Infrared (NIR) spectral bands, where water bodies exhibit high reflectance in the green band and strong absorption in the NIR band. The NDWI is computed using Eq. (1), where ρ_Green represents the reflectance of the green band and ρ_NIR represents the reflectance of the NIR band. NDWI has been widely applied in previous studies for effective detection of waterbodies, particularly lake boundaries. Özelkan et al. (2020) analyzed three NDWI band combinations—Green and NIR, Green and Short-Wave Infrared (SWIR) 1, and Green and SWIR 2—using Landsat-8 Operational Land Imager (OLI) imagery and reported that the Green-NIR combination yielded the highest accuracy in delineating lake boundaries. Similarly, Rokni et al. (2014) conducted a comparative analysis of water detection indices such as NDWI, Modified NDWI (MNDWI), and Automated Water Extraction Index (AWEI) using multi-temporal Landsat-5 Thematic Mapper, 7 Enhanced Thematic Mapper Plus, and 8 OLI images. They found that NDWI demonstrated superior performance in extracting water boundaries, especially in mountainous terrains and small to medium-sized lakes.

\(\begin{align}N D W I=\frac{\rho_{\text {Green }}-\rho_{N I R}}{\rho_{\text {Green }}+\rho_{N I R}}\end{align}\)       (1)

In this study, Band 2 of the PlanetScope imagery was designated as the green band, and Band 4 as the NIR band(Table 2) to calculate the NDWI for the study area. The PS2 instrument on the Dove Classic satellites (operational from July 2014 to April 29, 2022) captures four spectral bands—blue, green, red, and NIR—using the original PS2 telescope. Each scene covers approximately 25.0 km × 11.5 km. The PS2.SD instrument on the Dove-R satellites—an upgraded version of PS2 with improved Bayer pattern and passband filters—also captures the blue, green, red, and NIR bands and provides a larger scene size of about 25.0 km × 23.0 km; imagery is available from March 2019 to April 22, 2022. The PSB.SD instrument on the SuperDove satellites, launched in mid-March 2020 and currently in operation, uses a different optical system (the PSB telescope) and captures a broaderset of spectral bands: coastal blue, blue, green I, green, yellow, red, red edge, and NIR. Each PSB.SD scene covers approximately 32.5 km × 19.6 km. The spectral response of PSB.SD is designed to match that of PS2.SD for compatibility (Planet Labs PBC, 2025). The Green and NIR bands from PS2.SD and PSB.SD exhibit nearly identical wavelength ranges, whereas the original PS2 sensor shows a slightly broader and shifted spectral range. To mitigate potential discrepancies arising from these spectral differences, the NDWI threshold was not fixed but manually adjusted for each scene based on visual interpretation of water boundaries. This approach ensured that the NDWI-derived water surface area accurately reflected real conditions despite sensor variations. The workflow for calculating the surface area of impounded water in Baengnokdam is detailed in Fig. 2.

Table 2. Spectral band configurations of PlanetScope Dove satellite imagery, including PS2, PS2.SD, and PSB.SD products (Planet Labs PBC, 2025)

Fig. 2. Workflow for calculating the water surface area of Baengnokdam using PlanetScope imagery.

After calculating NDWI using PlanetScope satellite imagery, contour generation was performed using ArcGIS Pro software. Theoretically, NDWI values range between –1 and 1, and to enable precise extraction of water boundaries, the contour interval was set to 0.02. When contour intervals of 0.1 and 0.05 were applied, the resulting water boundaries were overly simplified and did not accurately reflect the actual shape of the waterbody. Therefore, for more refined boundary extraction, contours were generated at 0.02 intervals, and the value that most closely resembled the actual waterbody shape was selected for each image, based on the analyst’s visual interpretation, and then converted into polygons. In some images, small water pixels within the impounded water were misclassified as non-water. To address this, the Eliminate Polygon Part tool of ArcGIS Pro was used to remove polygons smaller than 100 m² that were located either within or adjacent to the waterbody. This process helped refine the water boundary and improve the accuracy of waterbody detection to better match the actual shape of the impounded water (Fig. 3).

Fig. 3. Removal of small polygon fragments after NDWI-based water area extraction: (a) before and (b) after applying the eliminate polygon part tool of ArcGIS Pro to remove polygon features smaller than 100 m².

2.2.2. Correlation Analysis between Rainfall and the Water Surface Area of Baengnokdam

To quantitatively evaluate the influence of rainfall on changes in the impounded water surface area of Baengnokdam, daily precipitation data were collected from the Witseoreum AWS station, located near Baengnokdam, for the study period covered by the satellite imagery. First, stationarity tests were conducted on both the rainfall and water surface area time series using the Augmented Dickey-Fuller (ADF) test (Dickey and Fuller, 1979). This test determines whether the statistical properties of a series, such as mean and variance, are constant over time, which is a key assumption in time series modeling. A p-value of less than 0.05 was interpreted as evidence that the time series is stationary. Next, autocorrelation analysis was performed for each variable to investigate the internal temporal dependencies within the data. Specifically, the Autocorrelation Function (ACF) and Partial ACF (PACF) were applied. The ACF plots reveal the strength of correlation between current values and their previous values at various lags, while PACF plots isolate the direct influence of a particular lag, controlling for the effect of intervening lags (Box et al., 2015). These analyses helped to assess the internal memory and lag structure of the time series, supporting the interpretation of correlation patterns.

To determine the time lag between rainfall events and the subsequent increase in water surface area, the Cross-Correlation Function (CCF) method was applied (Vrsalović et al., 2022). Time lags ranging from 1 to 7 days were tested, and correlation coefficients were calculated for each lag using the formula (Eq. 2). In Eq. (2), x represents daily precipitation, y represents the water surface area, and \(\begin{align}\bar {x}\end{align}\) and \(\begin{align}\bar {y}\end{align}\) are their respective mean values.

\(\begin{align}\operatorname{Correl}(x, y)=\frac{\sum(x-\bar{x})(y-\bar{y})}{\sqrt{\sum(x-\bar{x})^{2} \sum(y-\bar{y})^{2}}}\end{align}\)       (2)

Through correlation analysis, the correlation coefficients between rainfall and the water surface area were calculated for each time lag (in days), from the day of rainfall occurrence up to seven days after. Separate analyses were conducted for each year, and the years 2017 and 2020 were excluded due to insufficient water surface data, with fewer than 20 valid observations available. The time lag with the highest correlation coefficient was defined as the optimal lag time, interpreted as the time required for rainfall to effectively result in an increase in the impounded water surface area.

2.2.3. Correlation Analysis between Water Surface Area and Water Level in Baengnokdam

To derive the relationship between the water surface area and the water level of Baengnokdam, monthly minimum, maximum, and average water level data measured in the field for the years 2021, 2022, and 2024 were used (World Heritage Office, Jeju Special Self-Governing Province, 2025). Because long-term water resource monitoring and climate change assessments require estimating levels during periods without in-situ measurements, we modeled water level as a function of satellite-derived surface area for such periods. Water level data were available from April to September in 2021 and 2022, and from April to November in 2024. The year 2023 was excluded from the analysis due to the unavailability of water level data for that year.

Accordingly, regression analysis between water surface area and water level was conducted using data from the three years: 2021, 2022, and 2024. According to World Heritage Office, Jeju Special Self-Governing Province (2025), water level measurements were taken at the location marked with red points on the map shown in Fig. 1. For each year, monthly minimum, maximum, and average water level data were compared with the corresponding satellite imagery-derived water surface area values (minimum, maximum, and average) during the same period. If three or more satellite images were available within a given month, the minimum, maximum, and average water surface area values were calculated. However, if all images were acquired on the same day or on two consecutive days, and thus lacked temporal representativeness, only the average value was calculated, excluding the minimum and maximum. When only two images were available, the average was calculated; if only one image was available, that value was considered the monthly average.

Subsequently, a first-order linear regression equation (Eq. 3) was developed to describe the relationship between water level and water surface area. Additionally, a logarithmic regression equation (Eq. 4) was also computed for comparative analysis. In these equations, H represents water level (cm), A denotes water surface area (m²), α is the slope, and β is the intercept. To evaluate the explanatory power of the regression models, the coefficient of determination (R²) was calculated, enabling a quantitative assessment of the correlation between water level and water surface area.

H = αA + β       (3)

H = α ln(A) + β       (4)

3. Results and Discussion

3.1. Estimation of the Water Surface Area of Baengnokdam

3.1.1. NDWI Calculation Results for the Baengnokdam Area

When comparing the satellite imagery and NDWI results within the study area, a clear distinction between water bodies and land surfaces was observed, although seasonal differences in lighting conditions and optical properties of the waterbody may vary depending on the acquisition timing (Fig. 4). The NDWI values along the selected boundary of the Baengnokdam waterbody, derived from the full set of satellite images used in the analysis, exhibited slight variation depending on the acquisition date, with an average of –0.44 ± 0.10. Such variability in NDWI thresholds for multi-temporal waterbody monitoring is known to result from seasonal differences in lighting conditions and the varying compositions of the waterbody itself (Ji et al., 2009).

Fig. 4. Comparison of remote sensing imagery acquired on 20 October 2018: (a) PlanetScope true-color composite, (b) PlanetScope NDWI image, and (c) high-resolution aerial imagery. In (c), the yellow line indicates the waterbody polygon derived from the NDWI image, while the red line represents the waterbody polygon delineated from the aerial imagery.

To precisely evaluate the accuracy of the water-land boundary extraction, aerial imagery corresponding to the same date was obtained from the National Geographic Information Institute (https://www.ngii.go.kr/). This aerial imagery was georeferenced to match the PlanetScope satellite imagery, and the water boundary was manually digitized as polygons by visual interpretation (Fig. 4), serving as reference data for comparison with the NDWI-derived boundaries. The aerial imagery, with a spatial resolution of approximately 12 cm, offered higher precision than satellite imagery and was well-suited for detailed boundary delineation. In this process, the water surface area extracted from the NDWI-based analysis was 9,463.21 m², which was closely aligned with the water surface area derived from the aerial imagery (9,628.79 m²). The difference between the two measurements was 165.58 m², corresponding to an error of approximately -1.75% relative to the aerial imagery result. These findings suggest that the NDWI-based estimation of the Baengnokdam water surface area is reasonably accurate. The spatial discrepancy is attributable primarily to differences in spatial resolution between the PlanetScope imagery (e.g., 3.7 m) and the aerial imagery (12 cm), as well as to mixed-pixel effects along the land-water boundary, which can cause slight shifts in the delineated boundary.

3.1.2. Results of Water Surface Area Estimation in Baengnokdam

Analysis of the time-series data on the water surface area of Baengnokdam based on PlanetScope imagery acquired from April to November between 2017 and 2024 revealed seasonal variability, showing peak surface areas repeatedly in spring and early autumn (Fig. 5)—expanding in spring with thaw and heavy rainfall, peaking again in early autumn under monsoon/typhoon rains, and then gradually declining from mid-September to November.

Fig. 5. Time-series plots of daily precipitation (bars) and lake surface area (dots) in Baengnokdam from April to November for each year: (a) 2017, (b) 2018, (c) 2019, (d) 2020, (e) 2021, (f) 2022, (g) 2023, and (h) 2024.

The temperature data recorded at the Witseoreum AWS station are summarized in the box-and-whisker plots shown in Fig. 6. All three parameters—minimum, average, and maximum temperatures—exhibited a distinct increasing trend from March through May. In March, the average minimum temperature was –3.04 ± 0.64°C, with the distribution of minimum values remaining largely below the freezing point. In April, the average minimum temperature rose to 1.78 ± 0.79°C, marking a transitional period during which sub-zero temperatures shifted to above-freezing conditions. By May, it further increased to 6.18 ± 0.48°C, indicating a complete transition to consistently positive temperatures across all observations. This progressive warming trend suggests that the thawing process commenced in April and reached full development by May. In addition, spring rainfall contributed to the expansion of the water surface, with large surface areas repeatedly recorded across multiple years. According to the same meteorological data, May 2019 recorded 1,045.5 mm of monthly precipitation, which accounts for approximately 15.5% of that year’s total annual precipitation (6,725.0 mm). Similarly, May 2023 and May 2024 recorded 821 mm (18.8% of the annual total) and 930 mm (17.6% of the annual total), respectively. In May 2019 (Fig. 5c), May 2023 (Fig. 5g), and May 2024 (Fig. 5h), the water surface area of Baengnokdam exceeded 18,000 m², indicating that the peak surface area does not occur only in summer but also in spring. Given that the full water capacity of Baengnokdam is 20,912 m² (Ko, 2010), surface areas exceeding 18,000 m² can be considered extreme values.

Fig. 6. Box and whisker plots showing the temperature distributions for minimum (green), average (orange), and maximum (blue) values across March to May. Each box represents the interquartile range (25th to 75th percentile), and the whiskers extend to the minimum and maximum temperatures. The line inside each box indicates the median.

In September, surface area also reached annual peaks along with summer, likely due to continued seasonal rainfall patterns, including monsoons, typhoons, and heavy rain events extending from late August into early September. Specifically, in September 2019, 2020, 2021, and 2022, high monthly precipitation values were recorded as 1,631.5 mm, 1,627.5 mm, 1,277 mm, and 1,307 mm, respectively. Among these, August 2020 and 2021 also experienced high precipitation of 1,098 mm and 843 mm, respectively. These concentrated rainfalls likely had a direct impact on the expansion of Baengnokdam’s water surface. In early September 2020, under the influence of Typhoons Maysak and Haishen, the surface area peaked at approximately 21,512 m² (Fig. 5d), and in early September 2022, it reached about 19,827 m². However, from mid-September to November, a gradual decrease in surface area was generally observed, likely due to reduced rainfall and inflow during autumn.

3.2. Results of Correlation Analysis between Rainfall and Water Surface Area

To assess the temporal characteristics of daily rainfall and its influence on the water surface area of Baengnokdam, a series of time series analyses was conducted. The ADF test confirmed that both the daily rainfall and surface area time series were stationary at the 1% significance level, with test statistics of –11.886 and –6.945, respectively (p < 0.0001 for both). These results indicate that the statistical properties of the series, including mean and variance, remained constant over time, making them suitable for time series modeling without differencing. Autocorrelation analysis further revealed distinct temporal structures in the two variables. The ACF for rainfall exhibited statistically significant positive correlations up to lag 5, and the PACF showed sharp declines after lag 5. This pattern suggests short-memory behavior in rainfall data and supports the applicability of a low-order autoregressive model. In contrast, the surface area time series showed weak autocorrelation beyond lag 1, with both ACF and PACF plots indicating minimal temporal dependence. This implies that surface area variations are not strongly dependent on their own previous values but are likely influenced by external hydrological drivers.

To explore the lagged relationship between rainfall and water surface area, the CCF analysis was conducted for lag intervals ranging from the day of rainfall (lag day 0) to seven days afterward (lag day 7) (Fig. 7). The correlation on the same day (lag day 0) was slightly negative (r = –0.111), indicating that immediate changes in water surface area do not directly correspond to rainfall events. Among the analyzed lag intervals, the highest correlation was observed at a four-day lag (r = 0.496), suggesting a modest association whereby rainfall is reflected in increases in surface area about four days later. The peak correlation at a four-day lag may reflect delayed hillslope inputs and water-balance thresholds within the crater-rim catchment (McGuire and McDonnell, 2010). Rainfall likely infiltrates the permeable volcanic regolith and then moves downslope as subsurface throughflow, with travel times on the order of days before reaching the lake margin (Mosquera et al., 2020).

Fig. 7. Correlation between precipitation and water surface area of Baengnokdam across varying time lags.

3.3. Correlation Analysis Between Water Surface Area and Water Level

To quantitatively determine the relationship between the water surface area and water level of Baengnokdam, regression analyses were performed using monthly minimum, maximum, and average water levels along with satellite-derived surface area estimates from the years 2021, 2022, and 2024 (Fig. 8). When combining the data from 2021, 2022, and 2024, the linear model yielded R² = 0.741 (p < 0.001), while the logarithmic model yielded R² = 0.730 (p < 0.001). Both models demonstrated statistically significant predictive performance for water levels, though the linear model had a slightly higher explanatory power. The linear regression model produced a Root Mean Square Error (RMSE) of 43.43 cm and a Mean Absolute Error (MAE) of 34.54 cm, indicating that, when using monthly minimum, maximum, and average water levels, the predictive accuracy was limited relative to the total observed range of water levels. According to the linear model, a 1,000 m² increase in surface area is associated with an average water level rise of 24.1 cm, quantitatively representing the long-term hydrological response of Baengnokdam.

Fig. 8. Regression results showing the relationship between water level and lake surface area in Baengnokdam for the years 2021, 2022, and 2024. Panels (a) and (b) represent models based on the combined data.

This linear relationship between surface area and water level has also been demonstrated in previous studies. For instance, He et al. (2025) conducted a linear regression analysis using satellite imagery-based surface area data and in situ water level measurements for Dongting Lake in China, reporting high explanatory power with R² values up to 0.88 for monthly data. Similarly, Xu et al. (2020) estimated annual water levels for Lake Mead in the U.S. using ICESat-2 elevation data combined with Landsat-derived surface areas, achieving an R² of 1.00 and an RMSE of 1.06 m, indicating very high prediction accuracy. These examples demonstrate that linear regression is effective in quantifying surface area-water level relationships in lakes with diverse topographic and climatic conditions.

However, in this study, monthly minimum, maximum, and average water levels were used, which may introduce potential errors due to temporal mismatches between satellite image acquisition and water level measurements. These time gaps, especially during periods of high variability in rainfall, can lead to significant discrepancies in water level and reduce the accuracy of regression analysis. If exact water levels corresponding to the timing of satellite imagery were available, the accuracy of the surface area-water level correlation analysis could have been further improved.

4. Conclusions

This study quantitatively analyzed changes in the impounded water surface area of Baengnokdam, located at the summit of Mt. Hallasan on Jeju Island, using PlanetScope satellite imagery from 2017 to 2024, and examined its correlations with precipitation and water level. NDWI-based surface area estimates revealed clear seasonal and meteorological variability between April and November. Cross-correlation analysis showed that the water surface area responded most sensitively approximately four days after rainfall events, with a maximum correlation of r = 0.496. Regression analysis conducted for the years with available water level data (2021, 2022, and 2024) demonstrated a significant linear relationship between water surface area and water level, yielding an R² value of 0.741. These findings highlight the potential for estimating water level in Baengnokdam using satellite-based observations and demonstrate the feasibility of remote hydrological monitoring in inaccessible alpine environments. The results provide valuable baseline data for future efforts to predict changes in high-altitude waterbodies under climate change scenarios. They can support the development of ecological management and natural resource conservation policies in mountainous regions.

Author Contributions

Conceptualization: Hyun CU; Data curation: Kang M; Formal analysis: Kang M; Investigation: Kang M; Methodology: Kang M, Park J; Project administration: Hyun CU; Resources: Hyun CU; Software: Kang M, Park J; Supervision: Hyun CU; Validation: Kang M; Visualization: Kang M, Park J; Writing–original draft: Kang M; Writing–review & editing: All authors.

Conflicts of Interest

No potential conflict of interest relevant to this article was reported.

Funding

None.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.