Global hydrological cycle disturbances due to intensifying climate change have led to increased drought intensity and duration (Lee et al., 2021). The Korean Peninsula, in particular, has experienced accelerated spatial and temporal concentrations of precipitation, resulting in an increased frequency of localized droughts (Ahn et al., 2020). Drought, characterized as a slow-onset disaster persisting over months to years, presents challenges in defining clear onset and termination points and encompasses complex characteristics that propagate across broad spatial scales (Wilhite, 2000).

Research on drought early warning systems (DEWS) development including drought indices, has been conducted to mitigate drought damage. Drought indices serve as indicators for quantitative drought assessment and characteristic definition (Mishra and Singh, 2010). Most drought index studies focus on meteorological drought evaluation, directly expressing meteorological phenomena impacts through measures such as the standardized precipitation index (SPI). However, reported drought impacts correlate more closely with hydrological drought than meteorological drought (Sutanto et al., 2024). Hydrological drought, defined by available water resources including streamflow, reservoir storage, and groundwater, represents a significant reduction in all forms of water availability within the terrestrial phase of the hydrological cycle (Otache et al., 2020). The distinction between meteorological and hydrological drought is critical, as meteorological drought does not invariably propagate to hydrological drought. This discrepancy makes it challenging to directly apply meteorological drought indicators to groundwater or streamflow drought prediction (Nalbantis, 2008).

Effective drought planning and management from a water resource perspective necessitates developing hydrological drought early warning systems capable of providing timely information on short-term drought conditions (Sutanto et al., 2024). Hydrological drought indices include streamflow-based measures such as the streamflow drought index (SDI; Tareke and Awoke, 2022), standardized runoff index (SRI; Shukla and Wood, 2008), standardized stemflow index (SSI; Vicente-Serrano et al., 2012), and reservoir storage-based indices like the standardized reservoir supply index (SRSI; World Meteorological Organization and Global Water Partnership, 2016) and reservoir drought index (RDI; Nam et al., 2013). However, these indices require extensive streamflow and storage data and tend to focus on major reservoirs and rivers. This limitation constrains accurate drought status assessment and appropriate drought metric development for widely distributed small and medium-sized water bodies. Satellite image-based water surface area detection presents an effective alternative for water resource infrastructure management. Satellite images provide spatiotemporally continuous and objective data for broad-scale water body monitoring. Previous satellite imagery-based water body monitoring studies have demonstrated success in flood mapping and large-scale water body time series analysis (Wang et al., 2019; Zhang et al., 2020; Hamidi et al., 2023; Jamali et al., 2024). However, improvements remain necessary for integrated water resource management at administrative district levels. Optical imagery-based studies, despite high-resolution advantages, face limitations in time series data acquisition due to cloud and weather conditions (Prudente et al., 2020), with relatively limited research on integrated monitoring systems for diverse water bodies including small and medium-scale features.

This study proposes an integrated water body monitoring methodology at the administrative district level using Sentinel-1 synthetic aperture radar (SAR) images to overcome these methodological limitations. SAR images provide ideal data for long-term water body monitoring, unrestricted by weather conditions and diurnal cycles (Shen et al., 2022; Martinis et al., 2022). The study establishes an automated framework based on the Google Earth Engine (GEE) platform for satellite image collection, preprocessing, and analysis (Zhao et al., 2021), presenting a technical framework for efficient monitoring of temporal changes in all water bodies, including small reservoirs and streams. To validate the applicability of the proposed approach was evaluated by comparing water surface detection results with optical images-based normalized difference water index (NDWI) analysis for selected water bodies within administrative districts.

2.1. Study Area

This study employed the Sentinel-1 C-band SAR system as the primary observation platform for water body detection on the Korean Peninsula. The water bodies, including rivers and lakes, were selected as detection targets based on land cover maps from the environmental geographic information service (https://egis.me.go.kr/intro/land.do) to establish an integrated water body monitoring system at the administrative district level. The study areas were categorized into model training regions and validation regions for system verification. The training dataset utilized water body data constructed from Sentinel-1 satellite images at five major rivers (Namwon River, Hyeongsan River, Namcheon River, Nakdong River, and Taehwa River) as indicated in Fig. 1. These rivers are situated across diverse topographical conditions including mountainous, plain, and urban areas, with image patches extracted from selected points (river points) along each river.

Figure 1. Administrative districts and water distribution of three study areas in South Korea with water masks overlaid on Sentinel-2 true color composite images. The study areas include Gwangsan-gu of Gwangju Metropolitan City, Hapcheon-gun of Gyeongsangnam-do, and Ulju-gun of Ulsan Metropolitan City. The index map at the bottom left shows the geographical location of study areas and river dataset locations (Namwon, Hyeongsan, Namcheon, Nakdong, and Taehwa Rivers)

Ulju-gun, Hapcheon-gun, and Gwangsan-gu were selected as validation regions for empirical system verification. Fig. 1 illustrates the administrative boundaries and water distribution of the study areas - Gwangsan-gu of Gwangju Metropolitan City, Hapcheongun of Gyeongsangnam-do, and Ulju-gun of Ulsan Metropolitan City - overlaid with water masks from land cover maps on Sentinel-2 true color composite images. Gwangsan-gu features the Hwangnyong River and Yeongsan River traversing urban areas, exhibiting complex land cover characteristics with mixed urbanized and agricultural areas. Ulju-gun comprises a water network of major rivers including the Taehwa River and Hoeya River, along with multiple medium and small-scale reservoirs such as Saeyeon Reservoir and Daeam Reservoir. Hapcheon-gun contains the Hapcheon Lake, a large artificial reservoir within the Hwang River system, serving as a critical water resource management point.

2.2. Data

2.2.1. Sentinel-1 SAR Data

This study utilized Sentinel-1 C-band SAR system observation data as the primary input for developing a water body detection model. Sentinel-1, operated under the European Space Agency (ESA) Copernicus program, is a sensor system capable of allweather observation and stable surface characteristic acquisition (Torres et al., 2012). The analysis employed Level-1 ground range detected (GRD) products (10 m spatial resolution) acquired in interferometric wide (IW) mode.

The dataset consists of 1,043 river region subsets extracted from 151 original Sentinel-1 scenes collected between January 2018 and December 2023. Specifically, the time-series observation data includes Hyeongsan River (377 patches), Nakdong River (360 patches), Taehwa River (231 patches) Namwon River (48 patches), and Namcheon River (27 patches). All image data were collected through the Google Earth Engine platform, incorporating vertical-vertical (VV), vertical-horizontal (VH) polarization, and incidence angle information. Image acquisition and preprocessing were implemented as an automated process using Python-based GEE application programming interface (API). The preprocessing workflow included radiometric calibration using sensor calibration constants to derive normalized backscatter coefficients, geometric correction using the shuttle radar topography mission (SRTM) 30 m digital elevation model (DEM) for orthorectification, and terrain correction to address topographic distortions (Mullissa et al., 2021). For system validation, separate annual observation data of Sentinel-1 for 2023 were established for the regions of Gwangsan-gu, Ulju-gun, and Hapcheon-gun. The detailed methodology for administrative district-based data collection using GEE API is presented in Section 2.3.3.

2.2.2. Auxiliary Data

10 m resolution DEM data stored in GeoTIFF format and subdivision land cover map were utilized as auxiliary data to enhance water body detection accuracy. DEM elevation information serves as an essential element in identifying natural surface water flow paths and predicting potential water body distribution zones. This topographical analysis substantially contributes to improving the accuracy and reliability of satellite image-based water body detection (Du et al., 2018; Sun and Li., 2024). The land cover map was employed to define spatial distribution characteristics of water bodies and eliminate false detection areas in non-water body zones. Water-related features were extracted including (1) river features encompassing national and local rivers, (2) riverbank features including riverbanks and river peripheries, (3) water body features comprising agricultural reservoirs, multipurpose reservoirs, natural lakes, and artificial lakes, and (4) inland wetland features including riverside coastal wetlands and inland wetlands. As shown in Fig. 2, comprehensive water body monitoring was enabled by including not only permanent water bodies within the water system boundary but also potential water bodies with seasonal variability, thus reducing the risk of actual water body pixel elimination during masking procedures.

Figure 2. Spatial distribution of water-related features in a river system. (a) Overview showing river features, riverbank features, and inland wetland features extracted from land cover data. (b) Detailed view of the boxed area illustrating permanent and potential water areas overlaid on a high-resolution optical image.

2.3. Methodology

A semantic segmentation model based on the Swin transformer Large architecture was developed for water body detection using Sentinel-1 SAR images. The overall methodology consists of sequential processes including data preprocessing, model training and inference, and post-processing (Fig. 3a). For model development, a river water body dataset constructed from Sentinel-1 SAR images was utilized for training. The input data was structured as three channels, combining VV and VH polarization bands with 10m resolution DEM data to reflect topographical characteristics. A total of 1,043 images were constructed by dividing original Sentinel-1 images into 512 × 512 pixel patches, with the dataset having binary mask labels of water body regions for each image. 5-fold cross-validation was applied to evaluate the model generalization performance. Due to the characteristics of Sentinel-1 satellite images, patches extracted from a single original image captured during the same period possess similar characteristics. To prevent data leakage that could occur from the distribution of these patches across training, validation, and test sets, the division was performed by grouping at the original image level. From the complete dataset, 203 images (20%) were separated as the test set, and 5-fold cross-validation was applied to the remaining data. Due to the original image group-based division, the number of data points varies for each fold, with training data ranging from 687 to 716 images and validation data from 124 to 153 images. This group-based division approach ensures the independence of image patches acquired from the same region and period. Data augmentation was performed on the training set to increase data diversity and enhance model generalization performance (Ding et al., 2016; Furukawa, 2017). Considering the characteristics of SAR images and DEM data, geometric transformation-focused augmentation techniques were applied. These included combinations of random rotation (90 degrees), left/right flip, random size cropping within 350–450 pixels and resizing to original dimensions (512 × 512), and image translation within 10%. Additionally, image diversity was secured through fine adjustments of brightness and contrast within a 10% range. These augmentation techniques were applied to each original image, expanding the training data approximately threefold, with identical geometric transformations applied to label images to maintain alignment with input images.

Figure 3. Overview of the water body segmentation framework using Swin transformer. (a) Pipeline for SAR-based water body segmentation showing dataset preprocessing, cross-validation splits, data augmentation, and model training process. (b) Detailed architecture of Swin transformer backbone and UperNet decoder with Swin transformer block components.

2.3.1. Model Structure and Training Strategy

Water body detection from Sentinel-1 SAR images presents challenges due to the complex nature of SAR backscattering patterns and speckle noise (Pech-May et al., 2023). This study implements the Swin transformer model, which utilizes shifted windows for multi-scale feature extraction (Liu et al., 2021). The hierarchical feature pyramid structure of the model integrates local water body characteristics and global spatial patterns for accurate boundary delineation in water body segmentation tasks. The model is based on an encoder-decoder architecture, with learning parameters optimized for SAR image-based water body detection. The model structure visualization is presented in Fig. 3(b). The backbone network, the Swin transformer, divides input images into 4 × 4 patches and embeds them into 96-dimensional feature vectors. Statistical normalization is applied during this process to reflect the distribution characteristics of satellite imagery. Feature extraction proceeds through a four-stage hierarchical structure, with the initial stage processing at 128 × 128 resolution using a two-layer structure with three attention heads. Subsequent stages incorporate progressive resolution reduction and increased attention heads, notably extracting features with twelve attention heads at 32 × 32 resolution through six layers in the third stage. For computational efficiency, a 12 × 12 window-based self-attention mechanism was implemented, along with dropout (0.3) and normalization techniques (layer normalization, batch normalization) to prevent overfitting. The Swin transformer blocks alternate between window multi-head self-attention (W-MSA) and shifted window multi-head self-attention (SWMSA) to effectively extract local and global features. Each block incorporates layer normalization (LN) and multi-layer perceptron (MLP), with residual connections improving gradient flow. The decoder utilizes UPerHead for multi-scale feature integration. Feature maps extracted at each stage (channels: 96, 192, 384, 768) undergo processing at various pooling scales (1, 2, 3, 6), culminating in binary classification through integrated feature maps with 512 channels.

Model initialization employed transfer learning using pre-trained weights from the ImageNet-22K Swin-L model. The AdamW algorithm was adopted for learning optimization, with training stability ensured through a two-stage learning rate adjustment strategy and weight decay. CrossEntropyLoss served as the loss function, combining primary segmentation loss and auxiliary loss to enhance model performance.

2.3.2. Model Evaluation

For quantitative evaluation of the water body detection model performance, various assessment metrics commonly used in pixel-wise classification were employed. Each evaluation metric was derived from the confusion matrix, assessing the classification results of water body (positive class) and non-water body (negative class) pixels. Intersection over union (IoU), precision, recall, F1-Score, and accuracy were selected as primary evaluation metrics to conduct a multi-dimensional performance analysis of the model.

IoU evaluates the overlap between predicted and actual water body regions, offering the advantage of simultaneously considering both over-estimation and under-estimation. IoU is defined as:

Precision represents the proportion of actual water body pixels among those predicted as water bodies, focusing on the evaluation of false positives:

Recall indicates the proportion of correctly detected water body pixels among actual water body pixels, emphasizing the assessment of false negatives:

F1-Score, the harmonic mean of precision and recall, is utilized to evaluate the overall balanced performance of the model:

Finally, Accuracy represents the proportion of correctly classified pixels among all pixels, serving to evaluate the overall classification performance of the model:

2.3.3. Administrative District-based Water Body Detection System

Despite the superior characteristics of Sentinel-1 SAR images enabling weather-independent observation, operational limitations exist in large-area analysis due to individual image acquisition processes and complex preprocessing steps. Traditional SAR image acquisition methods require manual downloads and processing through specialized software such as sentinel application platform (SNAP), demanding significant user expertise and computational time (Song et al., 2021; Moskolaï et al., 2022). To address these limitations, a data acquisition automation system utilizing the GEE API has been developed. As illustrated in Fig. 5, the implementation of efficient data access and an automated preprocessed image download system through GEE API has optimized the data acquisition process, facilitating systematic data collection for nationwide water body monitoring.

Figure 5. Administrative region-based water body detection system. (a) Example of automated data acquisition process: from administrative boundary to merged Sentinel-1 patches through GEE API interface (exemplified with Ulju-gun region). (b) Flow diagram of the complete system architecture, including Sentinel-1 image download and water body detection process.

For efficient nationwide data acquisition, an administrative district-based data management and processing framework has been implemented. The administrative district-based image acquisition system, unlike conventional grid-based processing methods, ensures direct correlation with administrative management units. The system architecture comprises the following key components (Fig. 5): First, a systematic spatial division system based on administrative boundaries has been established. Administrative district spatial extents are formalized in GeoJSON format, and minimum bounding rectangles (MBRs) are generated to define data acquisition regions. Second, in the satellite image acquisition phase, Sentinel-1 images covering each administrative district within specified temporal ranges are identified and analyzed for spatial coverage. Minimum coverage ratio thresholds are established to ensure image validity, and multi-channel data incorporating VV and VH polarization data along with incidence angle information are acquired.

The post-processing phase encompasses quality validation and integration of acquired patch data, where patches are merged into single images while maintaining georeferencing information and ensuring continuity preservation between patch boundaries. The collected administrative district-level Sentinel-1 GRD Level-1 10 m images undergo preprocessing after stacking with DEM data as three bands, followed by water body detection through the implemented Swin transformer model. The post-processing stage includes the removal of false detection areas using land cover map-based water masks and the application of buffer zones (20 m) to river and reservoir areas, considering boundary uncertainties. The processed results are ultimately integrated at the administrative district level for systematic water body monitoring.

3.1. Model Training Results

The performance of the water body detection model was evaluated through 5-fold cross-validation. The cross-validation was conducted by dividing the complete dataset into five subsets, analyzing model stability and generalization performance through multiple evaluation metrics for each fold. Five evaluation metrics were utilized: IoU, precision, recall, F1 score, and accuracy. The results and averages of evaluation metrics for each fold are presented in Table 4.

Table 4 . Performance metrics of the water body detection model across 5-fold cross-validation including IoU, precision, recall, f1 score, and accuracy.

Fold N	Water IoU	Precision	Recall	F1-Score	Accuracy
Fold 0	0.9278	0.9613	0.9638	0.9626	0.9960
Fold 1	0.9268	0.9616	0.9616	0.9620	0.9959
Fold 2	0.9616	0.9592	0.9655	0.9623	0.9959
Fold 3	0.9273	0.9603	0.9643	0.9623	0.9959
Fold 4	0.9118	0.9761	0.9327	0.9539	0.9952
Avg.	0.9311	0.9637	0.9576	0.9606	0.9958

The IoU metric, quantifying the overlap between predicted and actual water body regions, achieved a mean value of 0.9311 (σ=0.0181). Cross-fold performance analysis revealed that Fold 2 demonstrated the highest IoU (0.9616) combined with balanced precision-recall metrics (0.9592 and 0.9655, respectively). Fold 4 exhibited a distinct performance pattern with the highest precision (0.9761) among all folds, but relatively lower IoU (0.9118) and recall (0.9327), indicating a tendency toward more conservative water body predictions. Despite these variations in individual metrics, the model maintained consistent overall accuracy across all folds (0.9952–0.9960), with standard deviations below 0.02 for all evaluation metrics, suggesting robust performance across different dataset configurations. The confusion matrix analysis for each fold is illustrated in Fig. 6.

Figure 6. Confusion matrices visualizing the binary classification performance across all five folds. Each matrix shows the percentage and absolute number (in parentheses) of True Positive, False Positive, False Negative, and True Negative predictions for water body detection.

Validation on the test dataset was conducted using the Fold 2 model, which exhibited optimal performance. The validation process incorporated comparison with NDWI (McFeeters, 1996) results derived from Sentinel-2 images of corresponding regions in 2018 and 2021. As shown in Fig. 7, the water body detection results from the Sentinel-1 SAR-based model demonstrated high spatial correspondence with NDWI analysis results. This suggests effective model learning of SAR backscattering characteristics for water body region identification with high accuracy. As evidenced in the third and fourth rows of Fig. 7, the model maintained consistent detection performance even under conditions where Sentinel-2-based water body observation was constrained by cloud cover, demonstrating the effective utilization of SAR all-weather observation capabilities.

Figure 7. Visual comparison of water body detection results. From left to right: Sentinel-1 SAR input images, ground truth masks, predictions from proposed SwinL model, Sentinel-2 true color images, and corresponding NDWI results. White pixels in second and third columns indicate water bodies, while blue (water) and green (non-water) represent NDWI classification results.

3.2. Application Results of Administrative District-Based Water Body Detection

To validate the practical applicability of the developed deep learning model, temporal water body distribution analysis was performed for three administrative districts: Gwangsan-gu in Gwangju Metropolitan City, Ulju-gun in Ulsan Metropolitan City, and Hapcheon-gun in Gyeongsangnam-do Province during 2023. For quantitative model evaluation, the detection results from selected regions of Sentinel-1 SAR images were compared with the NDWI derived from Sentinel-2 images. The analysis focused on specific water bodies within each administrative district: the Jijeong reservoir in Gwangsan-gu (effective storage capacity: 1,150,000 m³, catchment area: 740 ha, full water level: EL.26.81 m), Sayeon lake in Ulju-gun (effective storage capacity: 19,510,000 m³, catchment area: 124.5 km², full water level: 60 m), and the southern section of Hapcheon lake in Hapcheon-gun. The availability of Sentinel-2 images was limited due to cloudinduced data gaps and temporal resolution constraints. Therefore, correlation and binary agreement analyses were conducted only on image pairs with temporal gaps of no more than 10 days between the two satellites. The correlation analysis evaluated the relationship between predicted water body results and NDWI values, while the binary agreement analysis assessed the spatial concordance of water/non-water binary classification results. Detailed quantitative results of the temporal average correlation and binary agreement values for each region can be found in Table 5.

Table 5 . Statistical summary of water body detection results for selected water bodies in administrative districts.

Administrative districts	Study site	Number of observations	NDWI threshold	Correlation	Binary agreement	Water area ratio (%)
Gwangsan-gu	Jijeong Reservoir	8	-0.010 (-0.061–0.122)	0.783 (0.592–0.908)	0.988 (0.984–0.991)	10.7 (3.9–14.8)
Ulju-gun	Sayeon Lake	8	-0.030 (-0.079–0.023)	0.756 (0.665–0.858)	0.985 (0.979–0.991)	6.9 (4.9–10.3)
Hapcheon-gun	Hapcheon Lake	10	0.003 (-0.047–0.053)	0.863 (0.777–0.916)	0.972 (0.968–0.976)	23.5 (17.3–28.9)

The analysis of Gwangsan-gu (Fig. 8) demonstrated effective water body detection even in urbanized environments surrounding the Jijeong reservoir. The temporal analysis throughout 2023 captured distinct changes in water surface area between February 17, when it reached its maximum, and September 9, when it recorded its minimum. Comparison with Sentinel-2 NDWI analysis results (Fig. 8b) exhibited high spatial concordance with an average correlation of 0.783 and binary agreement of 0.988 during the analysis period. The NDWI threshold for this region had a mean value of –0.010 with a range from –0.061 to 0.122. These results indicated accurate detection of temporal changes in water surface area. The Ulju-gun analysis (Fig. 9) evaluated water body detection performance for Sayeon lake (effective storage capacity: 19,510,000 m³), which is significantly larger than the Jijeong Reservoir (effective storage capacity: 1,150,000 m³). Despite conditions prone to geometric distortion and shadowing in SAR images due to its location in mountainous terrain, the model effectively detected water bodies during both minimum water surface area on March 8, 2023 (correlation: 0.742, binary agreement: 0.991) and maximum on October 10, 2023 (correlation: 0.858, binary agreement: 0.979). The NDWI threshold for this region had a mean value of –0.030 with a range from –0.079 to 0.023. Analysis of eight temporal observations demonstrated detection accuracy with average correlation of 0.756 and binary agreement of 0.985. The Hapcheon analysis of the southern section of Hapcheon lake revealed changes in water body boundaries between February 12, 2023, when the minimum water surface area was observed (correlation: 0.777, binary agreement: 0.968), and December 9, 2023, when it reached its maximum (correlation: 0.857, binary agreement: 0.969). The NDWI threshold for this region had a mean value of –0.003 with a range from –0.047 to 0.053. These results showed high spatial concordance with NDWI results (average correlation: 0.863, binary agreement: 0.972 during the analysis period). Analysis across these three regions demonstrated that the developed model enables effective water body detection under diverse environmental conditions, including urban reservoirs, mountainous streams, and large-scale lakes. Comparisons with optical image-based NDWI analysis exhibited high spatial concordance across all regions (average correlation: 0.756–0.863, binary agreement: 0.972–0.988), indicating potential applicability as an integrated water body monitoring tool at the administrative district level.

Figure 8. Multi-temporal analysis of water body detection in Gwangju (Gwangsan-gu) during 2023. (a) Comparison of detected water bodies on Sentinel-1 and NDWI between minimum (2023.02.17) and maximum (2023.09.09) periods. (b) Time-series analysis of jijeong reservoir showing SAR backscatter, water detection model results, NDWI, and overlay analysis from January to December.

Figure 9. Multi-temporal analysis of water body detection in Ulsan (Ulju-gun) during 2023. (a) Comparison of detected water bodies on Sentinel-1 and NDWI between minimum (2023.03.08) and maximum (2023.10.10) periods. (b) Time-series analysis of a sayeon lake segment showing SAR backscatter, water detection model results, NDWI, and overlay analysis from January to December.

The technical performance and operational limitations of the water body detection system developed in this study were analyzed to identify practical applications in water resource management and future improvement directions. The administrative district-based processing framework achieved through the Swin transformer-based model offers practical advantages over grid-based approaches by enabling direct integration with existing water resource management units. While previous studies focused on improving technical accuracy of water body detection (Zhang et al., 2020; An and Rui, 2022), this framework enhances operational utility through direct integration with administrative water resource management systems. Integration with GEE significantly reduced processing time compared to traditional SNAP-based workflows, enabling near real-time monitoring capabilities. However, this approach requires careful consideration of administrative boundary effects when water bodies span multiple administrative districts.

Several technical limitations warrant consideration. First, although the model shows consistent performance across various environmental conditions, the detection of small water bodies presents detection limitations. Given the spatial resolution of Sentinel-1 SAR (10 m × 10 m = 100 m²/pixel) and backscattering characteristics, while water bodies smaller than 20,000 m² (2 ha, approximately 200 pixels) can be detected at a single time point, reliable monitoring of their temporal changes remains challenging. This limitation stems from the increased influence of speckle noise and surrounding terrain backscatter, and seasonal/weather-dependent variations on smaller water bodies, which affects the consistency of backscattering signals across different time periods. Therefore, additional validation is required for these smaller-scale water bodies. Second, masking procedures may limit the detection of rapid water body changes such as temporary flood inundation areas or river overflow during intense rainfall events. Third, monitoring temporal resolution is constrained by Sentinel-1A revisit periods, potentially limiting system application in rapid change detection scenarios. Nevertheless, the administrative district-based water body monitoring system proposed in this study has demonstrated practical value through effective integration with existing water resource management frameworks. Future enhancements in small water body detection capabilities and strengthened time series analysis are expected to increase the system utility.

This research presents an automated system for water body detection from Sentinel-1 SAR images utilizing a Swin transformer-based semantic segmentation model. The primary research outcomes are as follows:

1) The developed model exhibited robust performance in 5-fold cross-validation, achieving an average IoU of 0.9311 (σ=0.0181) and F1 score of 0.9606. The low standard deviations substantiate the model stability.

2) The implementation of a GEE API-based data acquisition automation system and administrative district-level processing framework addressed the complexities and temporal constraints inherent in conventional SAR image processing.

3) Validation studies conducted at three distinct sites - Jijeong Reservoir in Gwangsan-gu (Gwangju), Lake Sayeon in Ulju-gun (Ulsan), and the southern region of Lake Hapcheon in Hapcheon-gun - demonstrated high spatial correlation (correlation 0.756–0.863) with Sentinel-2 NDWI analysis across diverse environments.

The significance of this research lies in demonstrating the feasibility of an efficient water body monitoring system through automation of processes from preprocessing to deep learning-based detection, while leveraging the weather-independent characteristics of SAR images. Future research directions should focus on expanding system implementation to a national scale for comprehensive spatiotemporal distribution analysis of domestic water bodies and long-term time-series monitoring of major water systems. Such expansion would facilitate the identification of spatiotemporal variation patterns in domestic water resources influenced by climate change and anthropogenic factors, thereby providing empirical evidence for enhanced water resource management policy formulation.

Figure 4. Visualization of performance evaluation metrics for classification model validation. (a) Confusion matrix demonstrating the relationship between predicted and actual values through True Positive (TP), False Positive (FP), False Negative (FN), and True Negative (TN). (b) Intersection over Union (IoU) metric calculation.

Figure 10. Multi-temporal analysis of water body detection in Hapcheon-gun during 2023. (a) Comparison of detected water bodies on Sentinel-1 and NDWI between minimum (2023.02.12) and maximum (2023.12.09) periods. (b) Time-series analysis of a hapcheon lake south section showing SAR backscatter, water detection model results, NDWI, and overlay analysis from January to December.

Table 1 . Distribution of training datasets obtained from major river regions using Sentinel-1 SAR images.

River	Size	Number of patches	Acquisition period	Seasonal distributiona)
Hyeongsan	512, 512	377	2018.01–2023.12	104/78/108/87
Nakdong		360		88/88/96/88
Teahwa		231		64/52/51/64
Namwon		48		6/24/18/0
Namcheon		27		8/5/8/6
Total (512x512) 1,043

^a) spring/summer/fall/winter..

Table 2 . Summary of validation datasets from three administrative districts with Sentinel-1 SAR scenes and corresponding reference data in 2023.

Administrative district	Administrative district area (km2)	Analysis period	Number of scenes (n)	Reference data type
Gwangsan-gu	500.97	2023.01–2023.12	31	Sentinel-2 NDWI
Ulju-gun	1,062.83	2023.01–2023.12	31	Sentinel-2 NDWI
Hapcheon-gun	10,542.5	2023.01–2023.12	62	Sentinel-2 NDWI

Table 3 . Model specifications and hyperparameters used in water body segmentation with Swin Transformer.

Parameter	Value
Backbone	Swin-L Transformer
Window size	12 × 12
Embedding dimension (C)	192
Number of Heads	6, 12, 24, 48
Decoder	UPerHead
Input size	512 × 512
Batch size	2
Optimizer	AdamW
Base learning rate	1e-4
Weight decay	0.01
Loss function	CrossEntropyLoss
Model initialization	ImageNet-22K pre-trained weights
Learning rate schedule	Cosine decay (8,000 iterations)

This research was supported by Pukyong National University Development Project Research Fund (Philosopher of Next Generation, 2024).

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