Sub-Pixel Classification of Water Body from Sentinel-1 SAR Images Using Attention V-net

Abstract

As a result of the increasing frequency and intensity of hydrological disasters such as floods and droughts due to climate change, the importance of stable and precise water body detection technology is growing. Although Sentinel-1 Synthetic Aperture Radar (SAR) imagery, which can provide wide-area observation, has the advantage of all-weather usability, it has limitations in detecting small-scale rivers or complex boundaries due to its relatively low spatial resolution. To overcome these limitations, this study applied an Attention V-net-based Single Image-based Super-Resolution model to evaluate water body detection performance. The proposed model showed excellent performance on the test dataset, achieving an Intersection over Union (IoU) of 0.879 and an F1-score of 0.936. A qualitative analysis also confirmed its ability to maintain river continuity and precisely restore boundary areas. A regional analysis showed that while the detection of main rivers and tributaries was stable in the Russia dataset, the Houston dataset, which has a mix of urban and suburban areas, showed instances where roads or agricultural fields were incorrectly classified as water bodies. This suggests that the model is influenced not only by the SAR signal but also by local land cover characteristics. It is expected that if these limitations are supplemented in the future through multi-temporal SAR analysis, optical-radar fusion, and post-processing corrections, this study can contribute to real-world water resource management and flood monitoring.

1. Introduction

The intensification of climate change has led to a global increase in the frequency and intensity of hydrological disasters such as floods and droughts. Caretta et al. (2022) in the Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report Working Group II (AR6 WGII) explicitly state that climate change exacerbates the risks of extreme precipitation, floods, and droughts, which profoundly impact water resource systems and human societies. In recent decades, large-scale floods and droughts have indeed resulted in immense losses to social and economic sectors, including agriculture, industry, and infrastructure, in addition to human casualties. Against this backdrop, research on building various satellite-based datasets for precise monitoring of climate and environmental changes is actively underway (Yu et al., 2024; Gong et al., 2024). The accurate detection and mapping of water bodies—including rivers, reservoirs, and inundated areas—has emerged as an essential task for water resource management, disaster response, and environmental research. This is because precise identification of water body distribution and changes can contribute to predicting flood risk areas, securing water resources during droughts, managing ecosystems, and formulating climate change adaptation strategies.

Among the satellite remote sensing data used for water body detection, optical satellite images like Sentinel-2 provide high-resolution multispectral information, making them effective for distinguishing water bodies from other land covers. Indeed, Pekel et al. (2016) precisely mapped global surface water changes using Landsat and Sentinel-2 data, demonstrating the accuracy of optical image-based detection. However, a fundamental limitation of optical imagery is the restricted acquisition of images under cloudy or rainy conditions (Huang et al., 2018).

In contrast, Sentinel-1, equipped with Synthetic Aperture Radar (SAR), can observe under cloudy or rainy conditions, making it widely used for flood and water body monitoring(Amitrano et al., 2018). A case study in Korea has also reported the development of a GeoAI training dataset for water body detection in Cheonjiho Lake using TerraSAR-X, Sentinel-1, and ALOS-2 images (Lee et al., 2023). SAR images can be observed in all weather conditions, complementing the data gaps of optical images and playing a crucial role in emergency disaster situations.

Despite these advantages, Sentinel-1 SAR imagery has critical limitations for precise water body detection. Its relatively low spatial resolution and speckle noise make it difficult to detect small-scale water bodies such as narrow rivers, small reservoirs, or inundated urban areas. Furthermore, it can be confused with land covers that have similar backscattering characteristics, such as wetlands, shadow areas, and agricultural fields (McCormack et al., 2022). This issue is directly related to the frequent mixed-pixel phenomenon in remote sensing images, which blurs water body boundaries and degrades detection accuracy (Foody et al., 2005). While various studies have attempted to overcome these issues, research on applying super-resolution techniques to SAR-based water body detection remains in its early stages (Shen et al., 2020). Previous studies have often focused either on water body detection alone (Pekel et al., 2016; Amitrano et al., 2018; Lee et al., 2024) or on super-resolution techniques themselves(Foody et al., 2005; Yang et al., 2022; Zhang et al., 2023), without effectively integrating the two to solve the fundamental problem of boundary restoration in SAR imagery.

To address these multifaceted challenges, this study proposes a deep learning approach that directly integrates super-resolution into the water body detection framework. The proposed methodology is based on the Attention V-net architecture, which converts 256×256 Sentinel-1 input patches into 512 × 512 super-resolution water body detection results. This study conducted model training using only SAR data, without incorporating optical imagery as input. However, during the labeling process, Sentinel-2 images were used as an auxiliary reference to ensure the accuracy of the ground truth data, with human annotators manually identifying water bodies. This design allows the model to precisely reflect actual water body boundaries while accounting for the unique backscattering characteristics and mixed-pixel issues of SAR images.

The primary contribution of this research is to demonstrate how a super-resolution-based model can effectively overcome the inherent limitations of SAR data for hydrological monitoring. Our aim, therefore, extends beyond simple binary classification. We seek to validate the hypothesis that this integrated approach can precisely restore complex and ambiguous water body boundaries blurred by the mixed-pixel effect and maintain the spatial continuity of narrow hydrological features, such as small tributaries, which are often fragmented or missed entirely by conventional methods. The subsequent sections of this paper will present results that specifically highlight these super-resolution capabilities, providing both quantitative and qualitative evidence to validate the effectiveness of our proposed method in real-world scenarios.

2. Materials

2.1. Study Area

This study selected the Zeya River basin in Amur Oblast, Russia, and the Houston area in Texas, USA, as the study areas. Fig. 1 shows the study areas for the two regions, with (a) representing Houston, USA, and (b) and (c) representing the Zeya River basin in Russia.

Fig. 1. Study area: (a) Houston (descending, 2024-04-06), (b) Russia (ascending, 2019-06-28), and (c) Russia (descending, 2019-06-05). Basemap source: Earthstar Geographics.

The Zeya River is a major left tributary of the Amur River, with a length of approximately 1,242 km, a basin area of about 233,000 km^2_, and an average flow rate of 1,810–1,850 m³/s. The upper reaches of the Zeya River are home to the Zeya Nature Reserve, which is characterized by a taiga ecosystem with a distribution of spruce and oak trees. In addition, the large Zeya hydroelectric dam, completed in 1980, has significantly reduced the flood extent and duration downstream, leading to changes in floodplain wetlands and freshwater ecosystems. Remote sensing-based hydro-ecological studies have been conducted on this area(Nikitina et al., 2018).

The Houston area is a major city in southeastern Texas, USA, and is vulnerable to repeated flood damage due to extensive urbanization and its low-lying topographic features. In particular, Hurricane Harvey in 2017 was accompanied by record rainfall, causing floods of 1–2 meters deep throughout the city. This resulted in extensive inundation and socio-economic damage. In previous research, Houston is a representative case study area where SAR image-based flood detection and water depth estimation studies have been actively conducted, with flood depth calculation using Sentinel-1 and Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) data (Kundu et al., 2022) and urban inundation detection through Interferometric Synthetic Aperture Radar (InSAR) techniques (Chini et al., 2019) having been presented.

Therefore, these two sites were strategically selected to test the model against distinct, critical challenges. The Zeya River basin, with its complex network of natural tributaries and meandering channels, serves as an ideal testbed to evaluate the model’s ability to maintain spatial continuity and precisely restore fine-grained boundaries in a natural hydrological system. In contrast, the urban and suburban landscape of Houston, with numerous features like roads and building shadows that can be confused with water in SAR imagery, provides a rigorous test for the model’s robustness against false positives in an anthropogenic environment. By validating our model against these two key challenges, mapping fidelity in natural systems and discrimination accuracy in urban ones, this study provides a comprehensive assessment of the proposed super-resolution technique’s real-world applicability.

2.2. Study Data

The data for this study utilized Sentinel-1 satellite imagery. Sentinel-1 is equipped with a C-band sensor that enables all-weather observation and provides images with different spatial resolutions through its stripmap (SM) mode and interferometric wide swath (IW) mode. Table 1 summarizes the acquisition area, date, observation orbit, mode, polarization, and incidence angle information for the Sentinel-1 images used in this study. Considering these mode differences, this study utilized SM mode images for the training and validation datasets, and IW mode images for the inference dataset.

Table 1. Sentinel-1 images used in this study

For the Zeya River basin in Russia, SM mode images were acquired for the ascending orbit on June 28, 2019, and the descending orbit on June 5, 2019. For the Houston area in the United States, an SM mode image was acquired for the descending orbit on April 6, 2024. Among these, the ascending orbit image of the Zeya River basin and the descending orbit image of the Houston area were used as the model’s training dataset. Subsequently, the descending orbit image of the Zeya River basin was separately used as a test dataset for conducting qualitative and quantitative evaluations of the model.

Meanwhile, Sentinel-1’s IW mode images have the advantage of covering a relatively large area, but they have a lower spatial resolution compared to SM mode images. Considering this difference in resolution, this study did not use IW mode images during the training process but only during the inference stage. This was done to qualitatively analyze whether the proposed super-resolution water body detection model could restore detection results to the level of the SM mode from low-resolution inputs. In particular, since the IW mode is a primary observation mode with high practical utility, it was used to verify the effectiveness of the super-resolution technique. In conclusion, IW mode images from the Zeya River basin in Russia and the Houston area were applied to the inference dataset to evaluate the generalization capability and practical application potential of the proposed model.

This study also utilized both Sentinel-1’s VV and VH polarization data. VV polarization is advantageous for precise water body boundary detection as it clearly reflects the backscattering difference between water and land. VH polarization is known to be effective for inundation detection because it responds sensitively to water body areas even in regions with mixed vegetation or artificial structures (Alonso-Sarria et al., 2025). Due to these complementary characteristics, using VV and VH together has been reported to improve detection accuracy compared to using a single polarization (Sajid et al., 2025). Accordingly, this study aimed to enhance the stability and accuracy of the water body detection model by using both polarizations simultaneously.

3. Methods

Fig. 2 outlines the overall procedural workflow of this study, consisting of (a) SAR data preprocessing, (b) label data generation, (c) Artificial Intelligence (AI) model training and evaluation, and (d) IW mode inference.

Fig. 2. Workflow of this study: (a) SAR data process, (b) ground truth process, (c) model training and evaluation, and (d) IW inference.

First, we prepared suitable input images for analysis by performing preprocessing on Sentinel-1 SAR images, including decibel conversion, filtering, and orthorectification. Next, we determined the presence of water bodies by referencing Sentinel-2 images and created detailed, precise label data based on specific criteria. The constructed input and label data were then divided into patches and underwent a data augmentation process, considering class imbalance, to be used for training the Attention V-net-based super-resolution water body detection model.

The trained model’s performance was evaluated quantitatively and qualitatively using an independent test dataset. Finally, inference was performed on Sentinel-1 IW mode images to verify the proposed model’s ability to produce super-resolution detection results from low-resolution inputs.

3.1. SAR Preprocessing

This study utilized Sentinel-1 Single Look Complex (SLC) images as input data. For water body detection, the complex signals for each polarization (VV and VH) were first converted into intensity images (magnitude squared, |z|²) and then to a decibel (dB) scale. For water body detection, backscattering intensity is a more direct and stable indicator than the phase component. In fact, intensity-based approaches are reported to be the most intuitive and widely used methods in multi-temporal SAR change-detection studies (Baek and Jung, 2019). The dB conversion compresses the wide dynamic range of intensity values and homogenizes the backscattering distribution, which enhances the contrast between water and non-water bodies. This conversion simplifies the visual interpretation of the images and improves the stability of subsequent classification and detection algorithms(Amitrano et al., 2018).

Speckle noise, a unique characteristic of SAR images, is a type of multiplicative noise that not only degrades the visual quality of the image but also lowers the accuracy of classification and detection. Accordingly, this study applied a 3 × 3 median filter to the converted dB images. The median filter has the advantage of superior edge preservation compared to linear filters and can effectively suppress speckle noise while maintaining image boundary information when a small window size is used (Lee and Kim, 2022). This process can help increase the detection accuracy of fine topographic features, such as water body boundaries.

Following the median filter, orthorectification was performed for spatial geometric correction. First, a Lookup Table (LUT) was generated through Digital Elevation Model (DEM)-based orthorectification of the VH polarization image. Subsequently, the VV image, after undergoing intensity conversion, dB conversion, and median filtering, was transformed from the SAR coordinate system to a geographic coordinate system by applying the LUT generated from the VH image. This process ensures that both polarization images are aligned in the same coordinate system, enabling comparative and combined analysis between polarizations. Orthorectification is an essential step in SAR imagery as it removes geometric distortions and allows for analysis based on a geographic coordinate system (Small, 2011).

After orthorectification, adaptive interpolation based on pixel distance weighting was applied to correct for slight discrepancies in pixel spacing and resolution. This step ensures the homogeneity of spatial resolution and guarantees the consistency of input images for subsequent deep learning training. Finally, the preprocessed VV and VH images were converted to the GeoTIFF format and stored. This process minimizes physical and geometric distortions that can occur in SAR images and builds a standardized input dataset.

3.2. Label Data Generation

In this study, training label data for water body detection were constructed through visual interpretation and manual annotation of the Sentinel-1 dB images. We created polygon-shaped vector masks by directly distinguishing between water and non-water areas in a QGIS environment. While water bodies can be basically identified by the backscattering characteristics of SAR images alone, it can be difficult to make a determination in areas such as shadows, low-reflection terrains, or around structures. In such cases, we referred to Sentinel-2 multispectral images acquired with the minimum possible temporal gap from the Sentinel-1 data to ensure high label reliability. The temporal differences were minimal for most scenes, with examples including a 0-day gap for one Russia scene (S1/S2 on 2019-06-28), a 1-day gap for the Houston scene (S1 on 2024-04-06 vs. S2 on 2024-04-05), and a 3-day gap for the other Russia scene (S1 on 2019-06-05 vs. S2 on 2019-06-08). To ensure a completely cloud-free view of the Houston area, one additional reference image with an 11-day gap was utilized (S2 on 2024-03-26).

To ensure consistency among annotators during the labeling process, detailed annotation guidelines were established. These guidelines included specific methods for handling cases with ambiguous boundaries, such as bridges over water, riverbeds exposed by low water levels, and narrow river sections (Lee and Jung, 2023). These guidelines helped to reduce annotation uncertainty and enabled consistent and reliable labeling even under complex topographical conditions. The generated vector masks underwent a cross-validation process among the annotators to minimize errors. We focused on identifying and correcting sections with omissions or improperly handled boundaries. The final data were then converted to a raster format and used as label data for training the deep learning model.

3.3. Model Architecture

Fig. 3 illustrates the overall architecture of the proposed Attention V-net-based super-resolution water body detection model. The network is based on an encoder-decoder structure that is an extension of U-net. However, a key distinction of our proposed architecture lies in its primary function. While U-net is generally applied to segmentation tasks where the input and output share the same spatial resolution, our model is specifically designed for super-resolution segmentation. We therefore refer to our model as an Attention V-net to highlight this capability of generating a higher-resolution output 512 × 512 from a lower-resolution input 256 × 256.

Fig. 3. Attention V-net model architecture in this study.

The network consists of a multi-resolution encoder, a symmetric decoder, skip connections, and attention blocks. The encoder extracts low-level features through repeated convolution and pooling, while the decoder restores the resolution via upsampling, combining high-level semantic information. In this process, the skip connections directly transfer low-level features from the encoder to the decoder stage, which helps effectively preserve boundary and shape information (Dimitrovski et al., 2024). The attention blocks selectively emphasize important features in the channel and spatial dimensions, enabling precise discrimination of water body boundaries (Jiang et al., 2022; Jonnala et al., 2025). This allows the model to not only improve resolution but also suppress unnecessary noise and simultaneously restore the continuity and fine structure of water bodies.

3.4. Model Training

To create the dataset for model training, we used the preprocessed Sentinel-1 input images and the generated labels. First, we performed band-wise 0–1 normalization (min-max normalization) on the input images to ensure a consistent dynamic range of the observed values. Next, we applied a patch slicing method to convert the wide-area SAR images into efficient units for model training. The patch size was set to 512 × 512, and the patches were split with a 25% overlap in both the horizontal and vertical directions. This setting reduces the loss of boundary information and ensures consistency, while also mitigating the problem of small water bodies being missed by getting caught on the patch boundaries(Lee and Jung, 2023). The same slicing method was applied to the test dataset.

The data was split into training and test datasets using a scene-based division rather than a simple ratio split. Specifically, of the three total SM mode images, the ascending orbit image of the Russia Zeya River basin from June 28, 2019, and the descending orbit image of the Houston area from April 6, 2024, were used for the training dataset. The image of the Russia Zeya River basin from June 5, 2019, was used as an independent test dataset. By separating scenes with different orbital directions (ASC/DSC) and regions, we designed the study to reduce the bias that might arise from using only one area and to verify the model’s generalization performance under various observation conditions.

The problem of class imbalance is inherently present in water body detection. Since water body areas occupy a relatively small proportion of the total image, we applied oversampling and patch-based balancing techniques to prevent non-water areas from having an overly dominant influence during training. Furthermore, we performed data augmentation to ensure diversity in the training data and improve the model’s generalization performance. Specifically, we applied transformations such as rotation, horizontal and vertical flipping, and translation to account for various observation conditions.

This study aims to generate an output (a super-resolution water body map) with twice the resolution of the input image. Therefore, we downsampled only the input data to a size of 256 × 256. The label data were kept at their original resolution (512 × 512), configuring the dataset so the model could produce an output with double the resolution of the input.

Table 2 shows the main hyperparameters used for model training. The Adam optimizer was used, and a multiclass sparse Intersection over Union (IoU) loss function was applied to address class imbalance. The weights of all convolutional layers were initialized using a He normal initializer to ensure training stability and convergence speed. The learning rate was progressively changed from 0.005 to 0.00001, and the batch size was fixed at 30. The maximum number of epochs was set to 1,000, with early stopping applied to automatically terminate training if no performance improvement was observed for more than 100 epochs. These hyperparameter settings helped prevent unnecessary overfitting while achieving optimal performance.

Table 2. Hyperparameters of the trained Attention V-net model

3.5. Model Evaluation

Following model training, we evaluated its performance using a confusion matrix derived from comparing the model’s predictions against the ground truth from the test dataset. This matrix was used to calculate several standard performance metrics, including precision, recall, the F1-score, and IoU, to quantitatively assess the model’s detection accuracy and generalization performance.

First, precision signifies the proportion of pixels predicted as water bodies that are actually water bodies. It is defined by Eq. (1).

\(\begin{align}Precision = \frac{T P}{T P + F P} \end{align}\)       (1)

Here, TP denotes True Positive, and FP denotes False Positive. High precision indicates a low number of false detections and is beneficial for reducing the error of excessively detecting non-water areas.

Recall represents the proportion of actual water body pixels that the model correctly detected. It can be expressed as in Eq. (2).

\(\begin{align}Recall = \frac{T P}{T P + F N}\end{align}\)       (2)

Here, FN denotes False Negative. A high recall value means fewer water bodies were missed, making it a crucial metric for applications where omissions are critical, such as in flood and inundation detection.

Since precision and recall have a trade-off relationship, a Precision-Recall (PR) curve, which visually represents the relationship between the two metrics, was also analyzed. The PR curve shows the changes in precision and recall values at various thresholds and is known to be a more appropriate indicator of model performance than a Receiver Operating Characteristic (ROC) curve, especially in situations with imbalanced data (Saito and Rehmsmeier, 2015).

In addition, the F1-score was used to evaluate the balance between precision and recall with a single value. The F1-score is the harmonic mean of precision and recall, and it can be defined as in Eq. (3).

\(\begin{align}\text {F1-Score}= \frac{ 2\times \text {Precision} \times {\text {Recall}}}{\text {Precision + Recall}}\end{align}\)       (3)

Particularly in this study, where the proportion of the water body class is imbalanced, relying solely on accuracy might overestimate model performance. Therefore, the F1-score, which simultaneously reflects both precision and recall, can be a more reliable metric (Chicco and Jurman, 2020).

Lastly, IoU is a metric that evaluates the degree of overlap between the predicted water body area and the actual water body area. It is defined as in Eq. (4).

\(\begin{align}IoU = \frac{T P}{T P + F P + F N}\end{align}\)       (4)

IoU is the value of the intersection of the prediction and the ground truth divided by their union, representing the overall fit for a segmentation task. Since it considers both FP and FN simultaneously, it is widely used as a single metric and is considered the most standard performance measure in image segmentation research (Wang et al., 2022).

In addition to this quantitative analysis, we conducted a qualitative assessment. We applied the trained model to Sentinel-1 IW mode images to generate super-resolved water body maps. We then visually inspected these maps against high-resolution reference imagery to evaluate how well the technique restored detailed shapes and boundaries.

4. Results

4.1. Training Dataset Construction Results

Fig. 4 presents representative examples of the training dataset constructed in this study. Panels (a1–d1) and (a3–d3) show patches extracted from Sentinel-1 SM mode SAR images, which were converted into intensity images and scaled to dB. These images were also processed with filtering for speckle suppression and DEM-based orthorectification to ensure quality and geometric consistency. Panels (a2–d2) and (a4–d4) show the corresponding water body label masks, which were created based on visual interpretation with auxiliary reference to Sentinel-2 multispectral images. To maintain annotation consistency, separate detailed guidelines were applied to areas with ambiguous boundaries, such as bridges over water, riverbeds exposed by low water levels, and narrow river sections.

Fig. 4. Examples of AI data patches constructed through preprocessing and annotation. (a1–d1) and (a3–d3) show input patches derived from Sentinel-1 SAR complex images after intensity conversion and dB scaling, while (a2–d2) and (a4–d4) represent the corresponding water masks.

To ensure model robustness and generalization, the final dataset was constructed with several key characteristics. First, the class imbalance between water and non-water areas, a common issue in water body detection, was mitigated using a patch-based balancing strategy. Second, the dataset’s diversity was significantly enhanced through data augmentation, which included geometric transformations such as rotation, horizontal and vertical flipping, and translation. As a result, the constructed dataset provides a comprehensive basis for training, featuring a wide range of water body shapes and environmental conditions while minimizing the inherent noise and geometric distortions of SAR imagery.

4.2. Attention V-net Model Performance Evaluation

The model effectively restored the shape and boundaries of water bodies on the test dataset. Fig. 5 compares the actual SAR input images (a1–d1), label data (a2–d2), and model prediction results (a3–d3) for the Zeya River basin. For panels (a) and (b), the predictions for the main river sections were almost identical to the labels, and the boundaries were restored smoothly and continuously. The results clearly show that the water body naturally continued along the river’s flow without being disconnected, which is a significant advantage in applications where spatial continuity is important, such as flood or inundation detection. A tendency for blurry coastal areas in the input images, caused by mixed pixels, to be precisely reproduced through super-resolution segmentation was also confirmed.

Fig. 5. Model prediction results on the Zeya river test dataset. (a1–d1) Input Sentinel-1 SAR images, (a2–d2) reference water masks, and (a3–d3) predicted results. Panels (a–b) show well-predicted water bodies, while (c–d) illustrate cases of under-detection.

In addition, small islands within the water body or narrow land sections on the river were sometimes more finely distinguished than in the labels, demonstrating the model’s ability to precisely restore detailed structures beyond simple binary classification. The model’s predictions were stable in wide water areas, maintaining water body boundaries close to their actual width without excessive inflation. Artificial structures located over water bodies, such as bridges and levees, were accurately distinguished as non-water, and false detections in areas like roads or building rooftops were kept low.

Fig. 6 shows the PR curve. The mean Average Precision (AP) was approximately 0.995, with an AP of about 0.998 for the non-water class and 0.884 for the water body class. The high AP value for the non-water class is due to its large proportion within the scenes. Conversely, the relatively low AP for the water body class is a result of the low percentage of water bodies in the test dataset and confusion that occurs in complex boundary sections, such as narrow rivers and floodplains. Specifically, both false positives(misclassifications) and false negatives (omissions) were mainly observed in narrow river sections like small streams and tributaries.

Fig. 6. Precision-recall curve of the proposed Attention V-net model.

As seen in Table 3, the model’s recall was at a level of 0.932, which effectively suppresses the omission of actual water bodies, making it advantageous for water body monitoring applications. Precision reached 0.94, contributing to a reduction in unnecessary alerts by keeping the false detection of non-water bodies low. The F1-score was 0.936, confirming that the model effectively manages both false detections and omissions in water body detection. Since accuracy can be overestimated due to class imbalance, this study focuses on IoU and F1-score as key metrics. The IoU was 0.879, indicating a high degree of spatial agreement between the prediction and the label. This quantitatively supports the model’s excellent performance in restoring water body boundaries despite the boundary uncertainty caused by speckle and mixed pixels. However, some errors were still observed in complex environments, such as small rivers or localized inundation areas within urban centers.

Table 3. Performance metrics of the Attention V-net model

4.3. Inference Results on IW Data

To confirm the model’s generalization capability, additional inference was performed using IW mode data rather than SM mode data. Fig. 7 presents the results for three scenes in IW mode: the Houston area on April 1, 2024 (descending orbit); the Russia area on July 28, 2019 (ascending orbit); and the Russia area on June 6, 2019 (descending orbit).

Fig. 7. Inference results of the Attention V-net model applied to IW data: (a) Houston (descending, 2024-04-01), (b) Russia (ascending, 2019-07-28), and (c) Russia (descending, 2019-06-06). Columns represent (1) VH polarization, (2) VV polarization, and (3) model inference results.

IW mode has relatively challenging conditions for water body detection because of its lower spatial resolution and wider observation range compared to SM mode. Nevertheless, the model was able to stably extract major rivers and water bodies across the wide scenes, and the shapes of large-scale water bodies were restored relatively accurately without boundary distortion. In particular, in the Russian region, the flow of not only the main river but also the major tributaries was well-represented, confirming the model’s spatial generalization performance.

Fig. 8 presents cases where the inference was performed relatively well. Panels (a), (b), and (c) show scenes from the Houston (descending, 2024-04-01), Russia (ascending, 2019-07-28), and Russia (descending, 2019-06-06) IW mode orbits, respectively, with each column representing VH polarization, VV polarization, and the model’s prediction result. In all three regions, the main river and major water bodies were stably restored. Notably, the water body was detected continuously without interruption along the river’s flow. In the Houston area, not only large-scale rivers but also small water bodies like swimming pools in residential complexes were accurately detected, demonstrating the model’s spatial generalization ability to handle both wide river widths and small water areas simultaneously. In complex scenes with artificial structures like bridges and building rooftops, these objects were not misclassified as water bodies, confirming the model’s stable segmentation performance.

Fig. 8. Examples of well-predicted water bodies from IW inference results: (a) Houston (descending, 2024-04-01), (b) Russia (ascending, 2019-07-28), and (c) Russia (descending, 2019-06-06). Columns represent (1) VH polarization, (2) VV polarization, and (3) model inference results.

For the Russia region, which has a natural environment with few artificial structures, the boundaries of both the main river and its tributaries were maintained relatively smoothly. In particular, no misdetections of roads, bridges, or building rooftops as water bodies were observed. The water body boundaries were also seen to be continuously connected at the meandering parts of rivers and at the confluences of tributaries. This suggests that the super-resolution process was trained to preserve the structural continuity of rivers, not just to correct resolution. Additionally, detailed terrain features such as small islands and sandbars within the water bodies were also segmented with relative accuracy, showcasing the model’s ability for fine-grained restoration.

The magnified qualitative analysis of the water body boundaries, which provides explicit evidence for the sub-pixel classification approach, is presented in Fig. 9. This figure illustrates the model’s ability to resolve boundary ambiguity by visually comparing the original pixel scale with the super-resolution output, using zoomed-in sections from the well-predicted cases shown in Fig. 8. The low-resolution IW input images (panels a1, b1, and c1) clearly show that the sharp transitions between water and land are often represented by mixed pixels, resulting in blurry boundaries. This blurring highlights the inherent challenge of low spatial resolution.

Fig. 9. Qualitative analysis of the sub-pixel boundary restoration capability. Panels (a1, b1,c1) show magnified views of the Sentinel-1 IW SAR input images corresponding to Fig. 8. Panels (a2, b2,c2) show the corresponding 2× super-resolution classification results. (a) Houston (descending, 2024-04-01), (b) Russia (ascending, 2019-07-28), and (c) Russia (descending, 2019-06-06).

However, the proposed Attention V-net model successfully resolves this ambiguity by generating a sharp and continuous boundary at the 2× super-resolution level (panels a2, b2, and c2). The comparison between the original pixel size and the restored detail at the sub-pixel size confirms that the model is learning to extract and reconstruct the latent sub-pixel boundary information, rather than merely performing image interpolation. This consistent capability across various regions (Houston and Russia) is the key justification for our proposed sub-pixel classification methodology.

Fig. 10 presents examples where inference did not perform well. Similarly, (a), (b), and (c) show IW mode scenes from the Houston (descending, 2024-04-01), Russia (ascending, 2019-07-28), and Russia (descending, 2019-06-06) orbits, respectively, with each column showing the VH polarization, VV polarization, and the model’s prediction result.

Fig. 10. Examples of poorly-predicted water bodies from IW inference results: (a) Houston (descending, 2024-04-01), (b) Russia (ascending, 2019-07-28), and (c) Russia (descending, 2019-06-06). Columns represent (1) VH polarization, (2) VV polarization, and (3) model inference results.

In the Houston area, roads, artificial drainage ditches, and agricultural field sections were sometimes falsely detected as water bodies. This is because the complex land cover characteristics of urban and suburban environments can cause similar backscattering to water bodies in SAR imagery. Consequently, while the wide main river was restored with relative accuracy, the false detection rate tended to increase in the subdivided land cover areas within the city.

In the Russian region, most of the main river was detected stably. However, as shown in panel (b), some water bodies were omitted due to localized brightness inconsistencies in the VH polarization image. In addition, as seen in panel (c), disconnections and omissions were observed in fine tributaries that were only 2–3 pixels wide (approximately 20–30 m). This shows that even with the super-resolution process, complete restoration can be difficult due to the limitations of SAR spatial resolution and the mixed-pixel problem. These examples indicate that while the model shows stable performance in large water areas, detection limitations still exist in complex urban environments with artificial structures or in very narrow tributaries.

The comparison between Houston and Russia revealed that the model’s error patterns are closely related to regional characteristics. In the natural environment of Russia, the continuity of the main river and its tributaries was generally well-maintained. However, in urban environments with many artificial structures, such as Houston, misdetections of non-water areas occurred. This suggests that the model learns not only the backscattering characteristics of SAR but is also influenced by regional patterns present in the training data. This finding suggests that preprocessing or correction steps that consider the land cover characteristics of the target area are necessary for practical applications.

Due to the nature of IW data, which has a lower spatial resolution, small-scale rivers or canals less than 2–3 pixels wide (approximately 20–30 m) were often disconnected or omitted during the prediction process. This phenomenon of disconnection and omission of small rivers was also observed to some extent in SM mode, but the tendency was more pronounced in the lower-resolution IW mode. This indicates that even after undergoing the super-resolution process, it is difficult to completely overcome the limitations of the input data’s spatial resolution and the mixed-pixel problem. Nevertheless, in wide water bodies such as main rivers, lakes, and reservoirs, the water bodies were detected close to their actual boundaries without excessive expansion, which shows that the model maintains stable performance for large-scale water areas.

As a result, this study confirmed that the proposed Attention V-net-based model can effectively detect water bodies even in low-resolution data like IW mode, while reflecting diverse environmental characteristics. However, limitations such as confusion due to regional land cover characteristics (e.g., roads and agricultural fields in urban areas) and the omission of small rivers in low-resolution environments were also identified. Future research can consider overcoming these limitations by using time-series analysis of multi-temporal SAR images, complementing data with optical/radar fusion, or utilizing land cover masks in the post-processing stage. It is expected that this will further enhance the reliability of water body detection in diverse environments.

5. Discussion

This study evaluated the performance of an Attention V-net-based super-resolution model for water body detection by applying it to both SM and IW SAR images. A synthesis of quantitative metrics (IoU 0.879, F1-score 0.936) and qualitative results revealed that the model demonstrated stable performance across various observation modes and regional environments. Notably, it was confirmed that the model maintained the continuity of rivers without disconnections along their flow and, more precisely, restored blurry coastal areas caused by mixed pixels through the super-resolution process. This demonstrates the model’s strengths in boundary restoration and ensuring spatial continuity, going beyond simple pixel classification.

Differences related to regional characteristics were also evident. In the natural environment of the Russia region, which has few artificial structures, the detection of the main river and major tributaries was performed stably. However, in some sections of the main river, detection failed due to local brightness inconsistencies in the polarization images, and disconnections and omissions still occurred in narrow tributaries. In contrast, in the mixed urban and suburban environment of the Houston area, small water bodies like swimming pools in residential complexes, as well as the main river and tributaries, were detected with relative accuracy. However, areas like roads, artificial drainage ditches, and agricultural fields were sometimes misclassified as water bodies. This suggests that the model is heavily influenced not only by the SAR signal itself but also by regional land cover patterns.

We also analyzed the impact of different observation geometries by comparing the model’s performance on ascending and descending orbit data for the Russia region. A qualitative difference was noted, with the model generally performing better on the descending orbit images. Specifically, the ascending orbit image contained large areas of homogeneous low backscatter, resulting in poor contrast, whereas the descending orbit image exhibited a clearer contrast between the river and the surrounding land. This is likely attributable to the different acquisition dates and times, which lead to variations in environmental conditions such as soil moisture or surface roughness. These low-contrast areas in the ascending data presented a greater challenge for the model, occasionally leading to less confident boundary delineation. This finding suggests that while the model is robust, its performance is also highly sensitive to the quality and characteristics of the input SAR signal, which can vary significantly with observation time and environmental conditions.

While the model demonstrated strong performance, several limitations inherent to the methodology and data must be explicitly discussed. The foremost limitation is the inherent ambiguity of SAR backscatter signals, which resulted in both omissions and false positives. In the Russian region, localized brightness inconsistencies in polarization images led to the omission of actual water bodies.

Conversely, in the complex urban landscape of Houston, features with low backscatter signals, such as roads and agricultural fields, were frequently misclassified as water. This indicates that relying solely on SAR intensity is insufficient in diverse land cover environments. Furthermore, despite the super-resolution approach, the fundamental spatial resolution of the input data remains a constraint. As shown in the results, the model struggled to completely restore fine tributaries less than 2–3 pixels wide (approximately 20–30 m), as the necessary boundary information was absent from the original data due to the mixed-pixel effect. Finally, from an evaluation perspective, the relatively low AP for the water body class, despite high overall IoU and F1-scores, highlights the well-known challenge of model evaluation on highly imbalanced datasets, where performance on the rare positive class can be overestimated by traditional metrics.

Future research should incorporate time-series analysis of multi-temporal SAR images to reflect seasonal and environmental variability, and use fusion with optical imagery to mitigate confusion issues in areas like roads and agricultural fields. Additionally, applying a land cover mask or a connectivity correction algorithm in the post-processing stage could contribute to restoring disconnected tributaries. Furthermore, by applying next-generation high-resolution SAR data like NISAR, it is expected that we can overcome the limitations of small water body detection and build a highly reliable system that can be directly utilized for real-world water resource management and flood monitoring.

Author Contributions

Conceptualization: Jung JH; Data curation, Formal analysis: Jung JH; Investigation, Methodology: Jung JH, Yu JW; Project administration: Jung JH, Yu JW; Resources: Jung JH; Supervision: Jung HS; Validation: Jung JH, Yu JW; Visualization: Jung JH; Writing–original draft: Jung JH, Yu JW; Writing–review & editing: Jung JH, Jung HS.

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.