From Object-Based Classification Outputs to Training Labels:A Label-Efficient Pipeline for Water Body Segmentation

Abstract

Water body information is essential for water-resource management and water-related disaster response.However, building high-quality training data remains labor-intensive, motivating the development of methods that improve detection accuracy while reducing labeling effort. In this study, we propose a pipeline that automatically generates water labels from object-based image analysis (OBIA) classification outputs and applies them to the construction of a deep-learning dataset for water segmentation. Water pixels are first extracted from the OBIA class map. Next, the area of the largest connected water component is computed, and an image-specific adaptive threshold is applied to remove small false detections. Finally, morphological closing with different combinations of kernel size and iteration number is used to mitigate boundary noise. Using 40 images from the public OpenEarthMap (OEM) dataset, we compared the generated labels with manual ground truth (GT) using intersection over union (IoU). With an optimal closing configuration, the proposed refinement achieved a mean IoU of 0.827. After constructing deep-learning datasets, we trained a VGG19-based U-Net(VGG19-UNet) and evaluated the test F1-score (F1). Training with OBIA-derived labels yielded the highest mean F1 (0.882) and the lowest standard deviation (0.008) compared with training using manual or OEM labels, indicating improved stability and reproducibility.

1. Introduction

Water body information, including rivers and reservoirs, is widely used for flood and drought monitoring, water resource management, and detecting changes in aquatic environments. With the increasing availability of high-resolution remote sensing imagery, extracting water bodies while preserving fine-scale spatial patterns—such as riverbanks and reservoir embankments—has become increasingly important. However, automatic extraction from high-resolution imagery remains challenging. Water spectra can vary substantially due to shadows, dark roofs or roads, turbidity changes, and floating materials, while water–non-water boundaries are often complex. Although deep-learning-based semantic segmentation has substantially improved water mapping, its performance depends strongly on the quality of training labels. Constructing pixel-level labels manually for high-resolution imagery is time-consuming and costly, and maintaining annotation consistency is often difficult because delineation criteria may vary across annotators. Accordingly, label efficiency—constructing datasets that remain learnable with limited manual annotation—has emerged as an important research topic (Xia et al., 2023).

Index-based approaches that combine spectral bands with thresholding have long been used for water detection. Representative examples include the Normalized Difference Water Index (NDWI) (McFeeters, 1996), the Modified NDWI (MNDWI) (Xu, 2006), and the Automated Water Extraction Index (AWEI) (Feyisa et al., 2014), which enhance water bodies based on their spectral characteristics. These methods are simple to implement and scalable to large areas. However, in high-resolution urban scenes, shadows, roads, and building roofs often exhibit spectral responses similar to those of water, leading to false detections or omissions. Their performance is also sensitive to threshold selection, especially under strong illumination variation or heterogeneous land cover, where adaptive thresholding or clustering-based calibration is often required. At broader spatial and temporal scales, platform- and data-driven extensions have also been actively studied, such as global surface-water change products (Pekel et al., 2016) and large-scale processing on Google Earth Engine (GEE) (Gorelick et al., 2017). However, these studies are typically optimized for medium-to-low resolution data and are therefore less suitable for the fine-grained dataset construction targeted in this work.

In high-resolution imagery, pixel-based classification methods often suffer from noise and the lack of spatial context. To address these limitations, object-based image analysis (OBIA), also referred to as geographic OBIA (GEOBIA), has been developed (Blaschke, 2010; Hay & Castilla, 2008). OBIA groups pixels into meaningful objects and performs classification at the object level by jointly using spectral and spatial features. Object-based outputs often show improved boundary continuity and shape consistency, which is advantageous for GIS analysis and cartographic production. Previous studies have reported that object-based methods can outperform pixel-based methods in high-resolution urban land-cover classification (Myint et al., 2011). In particular, Ye (2021) proposed an HSV-based object classification approach that converts red-green-blue (RGB) imagery into the hue-saturation-value (HSV) color space, classifies segmented objects for each channel image, and integrates the results at the class level, achieving higher classification accuracy than commercial software.

Semantic segmentation has evolved substantially since the introduction of the Fully Convolutional Network (FCN) (Long et al., 2015) and U-Net (Ronneberger et al., 2015), and it has become a standard approach in remote sensing (Zhu et al., 2017). For water segmentation, networks and training strategies that address boundary complexity have been proposed continuously (Liu & Wang, 2022; Billson et al., 2023; Cao et al., 2024). Nevertheless, deep-learning models remain sensitive to label quality, particularly to boundary noise and detection consistency, making label generation and refinement important determinants of model generalization. In addition, weakly supervised, semi-supervised, and foundation-model-based approaches have been increasingly adopted to reduce labeling costs (Kirillov et al., 2023). From an operational perspective, it is important to automatically generate labels from existing products, such as OBIA classification outputs, validate their quality, and incorporate them into training workflows.

In this study, we propose a method for automatically generating water labels from OBIA classification outputs and using them to construct a deep-learning training dataset for water segmentation. The main contributions are as follows.

• We propose a pipeline for automatically generating water labels using OBIA classification outputs based on Ye (2021).

• We report intersection-over-union (IoU) statistics of OBIA-derived labels against manual ground truth (GT) and derive an optimal refinement setting for the experimental dataset.

• Using labels obtained with the optimal refinement setting, we train a VGG19-based U-Net (VGG19-UNet) model and discuss the performance and stability advantages of OBIA-derived labels.

2. Materials and Methods

2.1. Methodology

The proposed OBIA-based water-label generation pipeline consists of three stages: (1) generation of an OBIA classification output (class map); (2) generation of a binary water-label mask from the OBIA class map through adaptive threshold filtering and morphological closing; and (3) label-quality assessment and the construction of a deep-learning training dataset.

Fig. 1 summarizes the main procedure (water-candidate extraction →connected-component-based adaptive threshold Fig. 1 summarizes the main procedure (water-candidate filtering →morphological closing →quality logging). The input is the OBIA classification output produced following Ye (2021) (a color-coded class map). After generating a water-candidate mask M0, we compute the maximum connected-component area (A) using connected-component analysis. An adaptive threshold (T) is then applied to remove small false detections and generate a base mask M1. Morphological closing is applied to M1 to obtain the final mask M2. Along with the mask outputs, per-image IoU records and overall summary statistics are saved to ensure verifiability and reproducibility of the label-generation process.

Fig. 1. Flowchart of the proposed object-based image analysis(OBIA)-derived water label generation pipeline with adaptive thresholding and morphological closing, including quality-control records.

In this study, the OBIA result was generated based on the object-based classification framework of Ye (2021). However, the procedure was simplified for water-label generation. In Ye (2021), the input RGB image is transformed into the HSV color space. The hue component is then divided into multiple hue subchannels. Classification is performed independently for each subchannel, and class-label transformation and merging are subsequently applied to obtain the final object-based classification result. In the present study, however, we adopted a simplified strategy because water bodies generally exhibit smaller within-class color variation than buildings. Therefore, instead of dividing the hue component into multiple channels, we treated it as a single channel. Subclasses were constructed within each class only when necessary. The resulting classification outputs were then fused at the class level. Thus, the overall OBIA framework of Ye (2021) was retained in this study. However, the multi-hue-subchannel division step was omitted, and only the subclass construction and class-level merging strategy were applied.

For object generation, watershed-based image segmentation was applied to reflect the spatial continuity and boundary structure of image regions. To alleviate over-segmentation, a median filter was first applied, and low-gradient pixels were used as markers for watershed labeling. The classification result was then projected onto the segmented object regions. When more than one label was present within the same object, the final class was determined according to the dominant label ratio within that object. This object-level fusion reduced small isolated noise and unstable boundary classifications in high-resolution imagery, thereby producing a more spatially continuous and consistent initial OBIA output. This OBIA output was then used as the input for the post-processing pipeline shown in Fig. 1. Using this OBIA output as the input, a water-candidate mask M₀ was first generated, after which connected-component-based adaptive threshold filtering and morphological closing were sequentially applied.

2.1.1. Water Candidate Mask Extraction

The input is a color-coded OBIA classification map, and the water class is assumed to be encoded with a fixed color code. In implementation, we follow the OpenCV channel order of blue-green-red (BGR), and water pixels are represented as BGR = (255, 0, 0). The water-candidate mask M₀ is defined as follows:

M₀(x) = 1, if I(x) is c_water; otherwise, M₀(x) = 0       (1)

This step copies the pixels classified as water in the object-based output into the water-candidate mask without loss. That is, rather than binarizing continuous values using a threshold, the pipeline directly uses the discrete class decision from the OBIA output as input for label generation.

2.1.2. Adaptive Connected-Component Filtering

Because the water-candidate mask M₁ directly inherits the OBIA water classification, it may contain small false detections. To remove such artifacts, we compute connected components using an 8-neighborhood criterion and calculate the area of each component. Let A denote the area of the largest connected component. An image-specific adaptive threshold T is then defined as:

T = max(1, round(k · A))       (2)

Using this threshold, we construct a base mask M₁ composed only of connected components whose areas are greater than or equal to T. This strategy automatically adjusts the filtering strength in proportion to the representative water body scale A for each image, providing higher adaptability than an absolute area threshold when water body extent varies across images. Conversely, in images with a low water ratio, small water bodies may be removed; therefore, we analyze range-wise performance separately in the experimental results.

2.1.3. Morphological Refinement via Closing

To suppress noise near water boundaries, morphological closing is applied as a post-processing step:

M₂ = (M₁ ⊕ B) ⊖ B       (3)

where M₁ denotes the base mask after adaptive threshold filtering, M₂ denotes the final mask after closing, and B denotes the structuring element. Here, ⊕ and ⊖ represent morphological dilation and erosion, respectively. We compare six configurations by combining kernel sizes (3 or 5) with iteration numbers (1, 3, or 5). Here, Closing (a, b) denotes morphological closing performed with an a × a structuring element (kernel) for b iterations, where a indicates the kernel size and b indicates the number of iterations. Because excessive refinement may increase false positives (FP), we select the optimal configuration based on IoU against manual ground truth.

2.1.4. Quality Assessment and Output Products

Label quality is evaluated using IoU:

IoU = |P ∩ G|/|P ∪ G|       (4)

where P denotes the generated water label after morphological refinement, and G denotes the manual ground truth label. The F1-score is defined as

\(\begin{align}\text {F1-score} = \frac {2TP}{2TP + FP + FN}\end{align}\)       (5)

where TP, and FN denote true positives, and false negatives (FN), respectively.

To qualitatively analyze error patterns, we also generate comparison maps in which true positives (TP), true negatives (TN), FN, and FP are color-coded (TP = white, TN = black, FN = red, FP = green).

2.2. Materials

This study uses OpenEarthMap (OEM), a benchmark dataset widely adopted for high-resolution land-cover segmentation (Xia et al., 2023). OEM provides RGB imagery with a spatial resolution of approximately 0.25–0.5 m collected from multiple regions worldwide, along with eight-class labels that include the water class. Fig. 2 shows examples of OEM images and labels containing the water class. Fig. 2(b) illustrates a case where man-made structures present in a water scene are not properly represented in the label. Figs. 2(c, d) show examples where the water boundary is ambiguous, making it difficult even for manual annotation to delineate the water area accurately. In this study, we focus on the water class among the eight OEM classes and compare label-generation strategies and downstream training performance.

Fig. 2. Examples from the OEM water subset. The first row shows images, and the second row shows labels; white denotes water pixels. (a) A representative example of a typical water scene. (b) An example in which man-made structures are not properly represented in the label. (c, d) Examples in which the water boundary is ambiguous, making accurate delineation difficult.

Because some OEM water labels exhibited visually apparent mismatches with their corresponding RGB images, they were not adopted as the sole reference for quantitative evaluation. Fig. 3 presents three representative examples by comparing the RGB image, the OEM label, and the manually delineated label for each case. Among the 40 images examined in this study, two cases labeled artificial structures as water pixels, and one case incorrectly included vegetation pixels in the water class, resulting in an over-expanded water region. To avoid using such erroneous labels as the evaluation reference, a separate manually delineated GT was constructed. The GT in this study serves as a reference label for quantitative assessment and is not assumed to be a perfectly error-free absolute truth. Accordingly, the results should be interpreted as comparisons under the constructed reference GT, focusing on practical label quality and downstream learning effectiveness.

Fig. 3. Representative examples of apparent OEM water-label errors compared with RGB imagery and manually delineated labels:(a–c) Case 1, (d–f) Case 2, and (g–i) Case 3. For each case, the left, middle, and right panels show the RGB image, OEM label, and manually delineated label, respectively.

Table 1 summarizes the OEM dataset information used in this study. Among images containing the water class, we constructed two subsets by random sampling while considering the water ratio: (i) 1,000 images for deep-learning training and evaluation (OEM-1000) and (ii) 40 images for label-quality evaluation (OEM-40). The water ratio r is defined as:

\(\begin{align}r = \frac {|N_{water}|}{|N_{total}|} \times 100(\%)\end{align}\)       (5)

Table 1. OpenEarthMap dataset summary and subsets used in this study

where N_total is the total number of pixels in an image, and N_water is the number of pixels labeled as water.

To ensure that diverse water body patterns are represented in a balanced manner, all candidate images were divided into four bins according to r (Table 2). The deep-learning dataset (OEM-1000) was constructed by randomly selecting an equal number of images (250) from each bin, resulting in 1,000 images in total (Table 3). This stratified sampling is intended to include a balanced range of difficulties, from images with almost no water to images rich in water, and to mitigate bias that could occur if samples were concentrated in a single ratio range.

Table 2. Water-ratio bins used for stratified sampling

Table 3. Number of samples in each water-ratio bin for OEM-1000 and OEM-40

For label-quality comparison and validation of the OBIA label-generation pipeline, we additionally constructed an evaluation set of 40 images (OEM-40). Rather than focusing on Bin 1, where the water ratio is extremely low, OEM-40 emphasizes bins where water-object shapes are more clearly expressed (Bins 2–4): 10 images were randomly sampled from Bin 2, 20 from Bin 3, and 10 from Bin 4. For these 40 images, we prepared three label sources—the original OEM labels, labels generated by the proposed method (OBIA labels), and manual GT labels—and compared inter-label similarity using IoU.

2.3. Experimental Design

This section describes the experimental design based on the methodology and materials, including OBIA classification settings, parameter choices for water-label generation, quantitative evaluation protocols (IoU and F1-score), and the deep-learning training experiments. Section 3 then presents the quantitative results obtained under this design and discusses performance differences.

We generated OBIA classification outputs based on the hue-channel-based multi-class integration approach proposed by Ye (2021). Ye (2021) uses the hue component in HSV space to absorb intra-class color variation and presents a procedure to merge results at the class level. In this study, to simplify the workflow and reflect water body characteristics, we adopted the following simplified settings:

• The hue subchannel partitioning in Ye (2021) was not applied.

• Instead, subclasses were defined for each class to accommodate within-class variation.

• The subclasses were then merged into the base classes (class-wise merging) to generate the final class map.

• This simplification is based on the assumption that water bodies exhibit relatively low color diversity compared with buildings, such that subclass-to-class merging alone can provide sufficient benefit without subchannel partitioning.

In the OBIA output, water pixels are assumed to be encoded as RGB = (0, 0, 255) and are extracted as the candidate mask. To suppress noise, an image-specific adaptive threshold is applied, and k = 0.2 is used in the IoU evaluation. Because k determines the minimum connected-component area preserved in the base mask via T = round (k · A), an excessively small k may retain noise (small FP components), whereas an excessively large k may remove small water bodies and increase FN.

In our observations, k = 0.2 provided the best balance for this trade-off. For morphological refinement, we compared seven methods in total: the base water label before closing and six closing configurations (kernel sizes 3 × 3 or 5 × 5 with 1, 3, or 5 iterations). To justify the selected threshold setting, we first examined the effect of the adaptive threshold coefficient k before morphological refinement. As shown in Table 4, k = 0.2 achieved the highest mean IoU of 0.820, slightly outperforming k = 0.3 (0.817), whereas larger values (k ≥ 0.4) led to progressively lower performance. This trend suggests that excessively large thresholds may remove valid small water components, while overly small thresholds are less effective at suppressing minor noisy regions. Based on these results, k= 0.2 was selected as the default threshold coefficient for the subsequent refinement stage.

Table 4. Ablation results for the adaptive threshold coefficient k before morphological refinement on the OEM-40 dataset

After fixing k= 0.2, we further analyzed the effect of morphological refinement using different kernel sizes and iteration counts. As summarized in Table 5, Closing (3,3) and Closing (5,1) showed nearly identical mean IoU values when rounded to three decimal places. Using the full-precision results, however, Closing (3,3) achieved a slightly higher mean IoU (0.826706) than Closing (5,1) (0.826598). In contrast, more aggressive settings, such as Closing (5,5), consistently reduced the mean IoU. These results indicate that moderate morphological closing is effective for filling small gaps and smoothing object boundaries, whereas excessive closing may oversmooth narrow water structures and distort fine boundary details. Considering both performance and the use of a smaller kernel, Closing (3,3) was adopted as the final refinement setting in this study.

Table 5. Ablation results for morphological closing settings at k = 0.2 on the OEM-40 dataset

Overall, the ablation study confirms that the final parameter choice was not arbitrary. Rather, the combination of k = 0.2 and Closing (3,3) provided the best balance between removing small noisy components and preserving valid water regions, resulting in the highest overall label quality on the OEM-40 dataset.

To evaluate the quality of labels generated by the OBIA pipeline, we treated the manually produced labels for OEM-40 as the GT and computed the IoU of the OBIA-derived labels and the OEM-provided labels with respect to the GT. We report the mean, standard deviation, minimum, and maximum of IoU. In addition, we aggregated performance by water-ratio ranges (R0–9, R10–29, and R30–39) to analyze sensitivity to detection difficulty across images.

We used a common pool of 960 OEM images (and labels) as the baseline training set. From OEM-40, we fixed 10 images for validation and 10 for testing, keeping these splits identical across all experiments. The remaining 20 images were selected in a stratified manner to reflect the Bin 2/Bin 3/Bin 4 distribution (6/10/4), forming a stratified training subset (N = 20). During training, this subset was oversampled by a factor of five to increase its frequency, resulting in 1,060 training samples in total. To isolate the effect of label sources under controlled conditions, the labels of the 960-image training set were always fixed to the OEM labels, while only the labels of the stratified subset (N = 20) were replaced by OEM (baseline), OBIA, or manual (GT) labels.

To account for variability from random stratified selection, we repeated the construction with different random seeds and examined the stability of the results. The deep-learning model was VGG19-UNet, which combines the U-Net segmentation framework (Ronneberger et al., 2015) with a VGG19 encoder backbone (Simonyan & Zisserman, 2015). The model was trained for 100 epochs. Each experiment was run five times, and both the mean performance and variation were compared. Key hyperparameters and platform information are summarized in Table 6.

Table 6. Key parameters for deep-learning training and platform settings

3. Results and Discussion

This section presents quantitative results following the design in Section 2. We first compare IoU values across OBIA-label candidates and select the optimal refinement configuration, and then analyze range-wise performance characteristics of the selected setting. Next, we report IoU statistics for OEM-provided labels and discuss their characteristics relative to OBIA labels. Finally, we analyze downstream performance differences across label sources using deep-learning results (F1-score).

3.1. OBIA Label Similarity to Manual Ground Truth

Table 7 summarizes IoU statistics of seven OBIA-derived label variants generated with k = 0.2, evaluated against the manual GT over all 40 images (ALL: 0–39). Closing (3,3) achieved the best mean IoU, and we therefore adopted Closing (3,3) as the standard configuration in the subsequent dataset-construction step. Closing (5,1) followed closely. The kernel size of 3 with three iterations likely provided a balanced setting that improves boundary continuity and fills small holes within objects while suppressing excessive dilation. In contrast, Closing (5,5) showed the lowest mean and the largest standard deviation, suggesting that overly aggressive refinement can increase FP or distort structures, leading to larger performance variability.

Table 7. IoU statistics of OBIA-derived labels compared with manual ground truth (ALL set, k = 0.2)

Fig. 4 shows an example of water extraction using the OBIA classification output. In Fig. 4(b), some small regions in non-water areas are misclassified as water. After applying the size threshold, these small falsely detected regions are removed in the final output (Fig. 4c).

Fig. 4. Example of water extraction using the OBIA classification output: (a) RGB image, (b) OBIA classification output, and (c) extracted water mask after applying Closing (3,3) (white pixels).

Fig. 5 illustrates how the detected water mask varies under different closing configurations for the example shown in Fig. 4(c). As the kernel size and number of iterations increase, FN tend to decrease, while FP gradually increase.

Fig. 5. Visual comparison of the RGB image and water-detection results for different closing configurations: (a) RGB image, (b) OBIA result, (c) Closing (3,1), (d) Closing (3,3), (e) Closing (3,5), (f) Closing (5,1), (g) Closing (5,3), and (h) Closing (5,5).

Table 8 reports range-wise IoU statistics for Closing (3,3). The mean IoU is very high with small variance in R30–39, indicating that the pipeline operates stably on images where water bodies are large and clearly defined. In R0–9, the lower mean and larger variance suggest that small water bodies may be fragmented or have weak candidate connectivity during OBIA classification, and that the area-based filtering step can remove small water bodies, with limited recovery from closing alone.

Table 8. Range-wise IoU statistics of OBIA Closing (3,3) compared with manual ground truth (k = 0.2)

3.2. OEM Label Similarity to Manual Ground Truth

Table 9 presents IoU statistics of OEM-provided labels against the manual GT. As dataset-provided labels, OEM labels exhibit an overall agreement (ALL) similar to that of OBIA Closing (3,3). Specifically, the ALL mean IoU values for OEM (0.827) and OBIA Closing (3,3) (0.827) are identical. The higher OEM mean in R0–9 is consistent with the possibility that OEM labels capture small water bodies more frequently, whereas OBIA’s area-based filtering may remove small objects. In R10–29 and R30–39, OBIA Closing (3,3) is higher or comparable to OEM, suggesting that OBIA-based refinement is particularly effective in images with medium-to-high water ratios.

Table 9. IoU statistics of OEM labels compared with manual ground truth

3.3. Deep Learning Evaluation

Table 10 reports the test-set F1-scores and IoU values of the VGG19-UNet model after 100 epochs over five runs, while Table 11 summarizes the corresponding mean and standard deviation. Among the three label sources, the OBIA-based training setting achieved the highest mean F1-score (0.882) and the highest mean IoU (0.788). Moreover, it showed the lowest standard deviation for both F1-score (0.008) and IoU (0.012), indicating superior stability across repeated runs. This result suggests that the OBIA-derived labels provide a more stable representation of connected water structures, suppress small false detections through adaptive threshold filtering, and improve boundary continuity through Closing (3,3), thereby offering more favorable supervisory signals for model learning.

Table 10. F1-scores and IoU values of VGG19-UNet after 100 epochs on the test dataset across five runs

Table 11. Summary statistics of 100-epoch F1-scores and IoU values across five runs

From a stability perspective, OBIA achieved the lowest standard deviation. For large-scale dataset construction, repeatability of the overall cycle—label generation →training →performance—is important in addition to mean accuracy. The rule-based generation and standardized refinement of OBIA labels may reduce performance variability. A plausible explanation for the favorable performance of OBIA-based training is that labeling water regions is inherently difficult in high-resolution imagery. As shown in Fig. 3, visually apparent discrepancies can occur when water masks are affected by artificial structures or adjacent non-water regions. Moreover, water boundaries are often ambiguous because of mixed pixels, vegetation, shadows, and shoreline complexity, which may lead to fine-scale boundary variation in manually delineated labels. By contrast, the OBIA labels used in this study impose greater morphological consistency through area-based filtering and closing, which may produce labels that are more suitable for learning stable region-level patterns. Consequently, the model may become less sensitive to fine-scale variations that are not helpful for generalization, which may partly explain the higher test F1-score.

Fig. 6 compares model predictions across label sources. In cases with a low water ratio (first and second rows), all three label sources show a high proportion of FN. When water boundaries are ambiguous (fifth row), both FN and FP are high for all label sources. In the seventh row, OBIA-label training yields relatively fewer FN than the other two methods. For images with a high water ratio (rows 8–10), all methods achieve high detection accuracy.

Fig. 6. Comparison of deep-learning predictions on the test dataset: (a) RGB image, (b) manual GT, (c) prediction from training with manual labels, (d) prediction from training with OEM labels, and (e) prediction from training with OBIA labels.

4. Conclusions

This study presented a label-efficient dataset construction pipeline that automatically generates water labels from OBIA classification outputs. The proposed pipeline also jointly validates label quality and downstream deep-learning performance. In particular, we adopted a simplified strategy—building subclasses per class and applying only class-wise merging. This design improved practical usability and scalability while retaining the advantages of OBIA. On OEM-40, Closing (3,3) achieved the highest mean IoU (0.827) among the tested refinement configurations. Accordingly, it was adopted as the standard setting for subsequent dataset construction. In the 100-epoch deep-learning evaluation, training with OBIA-derived labels achieved a mean F1 of 0.882, outperforming manual and OEM labels. Moreover, it exhibited the lowest standard deviation, indicating robust repeatability across runs. These results suggest that OBIA-based automatic labels can reduce labeling costs while also improving training stability and generalization.

In future work, we plan to extend the OBIA-based automatic water-label generation pipeline—validated on the OEM benchmark—to optical imagery from KOMPSAT. This extension will enable evaluation of reproducibility and generalization under domestic operational satellite imaging conditions. The goal is to transfer the label-efficiency and learning-performance advantages observed on public benchmarks to operational satellite data. Ultimately, this transfer could reduce the cost of large-scale water-label construction while improving practical applicability.

Author Contributions

Conceptualization, Data curation, Funding acquisition, Investigation, Methodology, Project administration, Supervision, Validation, Writing–original draft, Writing–review & editing: Ye CS.

Conflicts of Interest

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

Funding

This work was supported by the 2025 Far East University Research Grant (FEU2025S04).

Data Availability Statement

The data used in this study are openly available from OpenEarthMap. The OpenEarthMap dataset can be accessed through the official project website and its public Zenodo repository at https://openearth-map.org and https://zenodo.org/records/7223446.