For segmentation experiments, we used five pan-sharpened WorldView-2 (WV2) images (red, green, blue (RGB) and infrared1 (IR1) bands only) with less than 20% cloud coverage, acquired between May and September in multiple years from 2010 to 2024. The imagery was provided by the Polar Geospatial Center (PGC) at the University of Minnesota through their cooperative agreement with the US National Science Foundations’ Office of Polar Programs. The spatial resolution of the WV2 images is approximately 0.5 m of ground sample distance (GSD), providing sufficient detail for land–water segmentation (European Space Agency, 2025). Each image is approximately 18,000–25,000 pixels in height and 17,000–25,000 pixels in width. The off-nadir angles for the five images were 11°, 25°, 31.6°, 42.9°, and 44.2°, reflecting differences in satellite viewing geometry which may influence shoreline and bluff visibility.

The geographic region was constrained to Elson Lagoon, a shallow embayment near the Iñupiat community of Utqiaġvik (Figure 1). Elson Lagoon was selected because Arctic Lagoons are geomorphologically diverse and ecologically significant, representing approximately 44% of the Alaskan Beaufort Sea coastline and offering a strong analog for Arctic coastal systems more broadly (Miller et al., 2021). Additionally, coastal erosion has been monitored along the Elson Lagoon shoreline for over 8 decades, providing a valuable resource of in situ data that can be used to validate newly developed machine learning-based approaches. The imagery captures diverse geomorphic elements, land-water interfaces, coastal bluffs, and permafrost-related landscape features, which are crucial for studying the impacts of climate change and coastal erosion in the Arctic. The dataset also reflects the unique imaging challenges of the region, including low contrast, mixed spectral textures between the shoreline, vegetation, and lagoon water, as well as seasonal variability. These complexities make the data particularly valuable for the Arctic more broadly, as they represent common issues encountered in Arctic satellite image analysis.

The five WV2 scenes from Elson Lagoon were manually annotated using ArcGIS Pro (version 3.5) to delineate land and water regions. These annotations served as ground-truth labels for the supervised model training. Annotation was performed by an expert with direct field experience at the site and prior in situ data collection over multiple seasons. The imagery selected for annotation featured ideal environmental conditions: minimal cloud cover, no snow or ice, clear visibility of land–water transitions, and favorable sensor viewing angles. The labeling scheme focused on identifying only land and water classes. Areas not belonging to either category, such as beaches, sloping bluffs, tundra surfaces, or small inland water bodies, were left unannotated and thus implicitly treated as background. This structure allows for clear delineation of the main surface types while preserving the natural transitions needed for subsequent boundary extraction.

Annotations were created as polygonal masks, enabling a consistent delineation of spatial regions. Polygon boundaries were drawn along visually identifiable transitions based on surface color, texture, and shading. In areas with low contrast or shadowing, expert interpretation was used, guided by site-specific knowledge and prior field observations, to approximate the most probable land–water separation. Despite favorable conditions, challenges such as spectral confusion between shallow water and wet sediment, vegetation gradients along bluffs, and surface moisture occasionally complicated boundary placement. To ensure consistency, detailed labeling guidelines were developed and applied uniformly in all five scenes, each requiring approximately 2.5–3.5 h to annotate. Figures 2A,B shows examples of the imagery and corresponding annotated polygons, where land and water features are visualized in tan and teal, respectively. Supplementary Figure S1 provides additional visualization of the manual annotations overlaid on WV2 imagery, illustrating the spatial characteristics and delineation detail of the digitized features.

The satellite imagery used in this study underwent multiple pre-processing steps to ensure compatibility with the U-Net and DifFeat models. These steps standardized spatial alignment, improved feature separability, and prepared the data for model input. The main steps are described below.

For all experiments, the dataset consisted of five WorldView-2 (WV2) satellite images acquired between 2010 and 2024. These images were divided into distinct subsets: three images were designated for training, one for validation, and one for testing. All tiles used for model evaluation were exclusively obtained from the held-out evaluation image, ensuring strict separation between training, validation, and test data. The tiling process resulted in 528 training tiles, 196 validation tiles, and 255 testing tiles, each measuring 512 × 512 pixels. While some tiles partially overlapped spatially across acquisitions, differences in capture year, illumination, and viewing geometry ensured that the evaluation imagery remained distinct from the training and validation datasets. This protocol was consistently applied to both the U-Net and DifFeat models, providing a robust basis for comparison and minimizing the risk of overfitting to specific acquisition conditions.

To assess model generalization beyond training conditions, we evaluated both U-Net and DifFeat on large-scale Arctic imagery: (1) a large-scale WorldView-2 (WV2) satellite scene ( 18,000 × 20,000 pixels) covering the Elson Lagoon region, and (2) a high-resolution unoccupied aerial vehicle (UAV) orthomosaic ( 20,000 × 20,000 pixels at 2 cm/pixel) from the same coastal area. These scenes were not part of the training or validation data, allowing us to evaluate cross-scale and cross-source generalization. The large-scale WV2 image was processed in the same manner as the training dataset, including band selection, normalization, and tiling into 512 × 512 pixel patches for model inference. The predictions of each tile were mosaicked to produce a seamless segmentation of the entire scene, with boundary extraction performed as described in Sections 2.6.1 and 2.7.1.

The UAV imagery was collected on 6 August 2024 using a Quantum Systems Trinity F90+ fixed-wing UAV equipped with a Micasense Altum-PT multispectral sensor. The flight was carried out 120 m above ground level (AGL) over the Elson Lagoon coastline, following a corridor mapping path with 80% forward and side overlap. The resulting dataset was processed using a standard structure-from-motion (SfM) workflow, producing a 2 cm/pixel ground sampling distance (GSD) orthomosaic. The images were downsampled by extracting 5120 × 5120 slices and resizing them to 512 × 512 to match the input size of the model, thus preserving large-scale spatial patterns while reducing fine spatial detail. The same preprocessing, inference, and boundary extraction steps as with the training data were applied to ensure methodological consistency. No model re-training or fine-tuning was performed for these tests. The results of the large-scale WV2 image and the UAV orthomosaic were used exclusively to assess the ability of the trained models to generalize to new spatial scales and imaging platforms.

U-Net, a fully convolutional neural network originally developed for biomedical image segmentation (Ronneberger et al., 2015), features an encoder–decoder structure with skip connections that allow simultaneous capture of global context and fine spatial detail. The encoder path extracts high-level contextual features through successive convolution and down-sampling operations, while the decoder path progressively reconstructs spatial details using up-sampling and convolution layers. Skip connections between corresponding encoder and decoder stages allow spatial information lost during down-sampling to be preserved and integrated during reconstruction, giving U-Net its distinctive U-shaped architecture.

In this study, U-Net was employed to perform semantic segmentation of land and water regions in high-resolution Arctic satellite imagery. Accurate delineation of these classes is essential for subsequent extraction of shoreline and bluff edge boundaries via post-processing. Because Arctic coastal transitions can be narrow, low contrast, and highly variable in appearance, the localization capability of U-Net is particularly well suited to capture fine-scale land-water interfaces that define these geomorphic boundaries. Training images and their corresponding label masks were prepared following the procedures described in Section 2.3, including normalization, tiling, and class balance adjustments, emphasizing tiles containing both land and water pixels.

To enhance generalization, data augmentation was applied during training using random rotations and horizontal and vertical flips, exposing the model to varied orientations of coastal features. The network was trained using a hybrid loss function combining Dice loss, which mitigates class imbalance, and Boundary loss, which sharpens land–water transitions by penalizing errors near class boundaries (Aryal et al., 2021). Optimization was performed with the Adam optimizer using a batch size of 8 and an initial learning rate of 0.001. A learning rate scheduler reduced the learning rate after 30 epochs to improve convergence stability. Model performance was evaluated using Intersection over Union (IoU), Precision, and Recall, computed for the land and water classes. These metrics provide a comprehensive assessment of segmentation accuracy and boundary fidelity, directly influencing the quality of the derived shoreline and bluff edge boundaries. Quantitative results are presented in the Results section.

Differentiable Feature Clustering (DifFeat) is an unsupervised learning technique that leverages feature representation learning for image segmentation (Kim et al., 2020). Its key advantage lies in its ability to perform segmentation without the need for labeled data, making it particularly valuable in Arctic settings where annotated datasets are scarce or expensive to obtain. DifFeat clusters pixels based on learned feature representations extracted from a convolutional feature map, where each pixel is represented as a multidimensional vector across channels. Clustering is performed by assigning each pixel to the channel (cluster) with the highest activation, effectively grouping pixels with similar spectral–textural characteristics. In the implementation, the initial and minimum number of clusters are user-defined parameters. The model begins from the specified maximum and iteratively merges similar pixel groups during optimization until the minimum threshold is reached, controlling the final segmentation granularity. The optimization follows the loss formulation of Kim et al. (2020), combining Cross-Entropy loss which encourages grouping of similar pixels and L1 loss which enforces spatial continuity and smoothness by penalizing abrupt feature changes between neighboring pixels. Together, these mechanisms produce spatially coherent clusters that correspond to distinct geomorphic surface types.

As an unsupervised approach, DifFeat generates multiple clusters from the input image slices (Figure 2C). The model does not assign semantic classes to these clusters and the model’s role is limited to grouping pixels with similar features. In our workflow, this is adapted into a minimally supervised method: a single manual step is used to visually inspect clusters from a reference slice and identify those corresponding to the two target classes (water and land). Once the target clusters are identified, their cluster IDs are fixed and used consistently across all slices and images, enabling fully automated extraction in subsequent predictions. If a selected cluster does not appear in a slice, it simply reflects that the class is absent from that portion of the scene. This fixed-ID mapping is central to reusing the model output without reintroducing manual work.

DifFeat’s segmentation approach provides flexibility in cluster granularity, supporting both coarse-grained and fine-grained clustering depending on the desired level of feature separation. An important advantage of DifFeat in our research is its ability to operate effectively with minimal data, processing as little as a single image slice of size 224 × 224 or 512 × 512 pixels. This is particularly useful in Arctic coastal monitoring, where data are scarce due to remote site locations and environmental challenges such as persistent cloud cover. The input image slice is obtained from the slicing process explained in Section 2.3.3, normalized to standardize the pixel intensities and improve the stability of the feature clustering. The model can process various combinations of spectral bands, including RGB, RGB+IR, or RGB+IR+NDWI+NDVI, providing flexibility in using additional information for clustering. The model is initialized with random weights and trained from scratch, ensuring that the learned features are specific to Arctic coastal imagery.

The resulting class-coded masks, which comprise land, water and background regions, were subsequently processed using the automated interface extraction workflow (Section 2.8) to derive boundaries of the shoreline and the bluff edge. This unified post-processing step ensures methodological consistency between the supervised and unsupervised approaches, enabling direct comparison of segmentation accuracy, generalization, and data efficiency. Figure 2 illustrates the complete DifFeat workflow: (A) input WV2 image slice, (B) ground-truth mask displayed for comparison only (not used for model training), (C) clustering or segmentation map obtained from the model, (D) automatically extracted shoreline and bluff edge boundaries from land and water segments, and (E) the boundaries overlaid on the input image. Together, these panels demonstrate how DifFeat, despite operating without labeled data, produces coherent land and water delineations from which geomorphically meaningful shoreline and bluff edge boundaries can be objectively derived.

Following segmentation, we derived geomorphic boundaries corresponding to the shoreline and bluff edge from the predicted land, water and background masks. Specifically, the model predicts two explicit classes, land and water, while unclassified regions (e.g., beach or bluff slopes) are treated as background. These boundaries are not directly predicted by the model but are instead derived from the spatial interfaces between the segmented regions. Boundaries were extracted by identifying adjacency relationships among these classes. For each class map, we first defined binary masks for land, water, and background. Using 8-connected morphological dilation with a 3 × 3 structuring element, we delineated geomorphic boundaries by locating water pixels adjacent to land or background (shoreline) and land pixels adjacent to water or background (bluff edge). To remove artifacts, small closed loops were filtered out and boundary pixels along image edges were discarded. This approach ensures that both shoreline and bluff edge lines are objectively derived from the predicted land and water regions, without requiring manual delineation for each image or direct training of linear features. The same procedure can be applied to either ground-truth or predicted masks, enabling consistent boundary extraction for evaluation and analysis.

Model performance was quantitatively assessed using four standard region-based metrics: Intersection over Union (IoU), Precision, Recall, and F1 score. All metrics were computed separately for the land and water classes using the manually annotated polygons as ground-truth reference. IoU was calculated as the ratio of the intersection to the union of predicted and ground-truth pixels for each class. Precision represents the proportion of predicted pixels that are correct, while Recall quantifies the proportion of ground-truth pixels that are correctly identified. The F1 score is the harmonic mean of Precision and Recall, summarizing segmentation accuracy for each class. These metrics were reported for both U-Net and DifFeat on the held-out evaluation set, allowing for direct comparison of segmentation performance across model types and spectral input combinations.

Table 1 summarizes the segmentation performance of U-Net and DifFeat in three spectral input combinations (RGB; RGB+IR; RGB+IR+NDWI+NDVI).

The training times, parameters, and configurations for the U-Net and DifFeat clustering models are summarized in Table 2. U-Net contains approximately 1.943 M learnable parameters, compared to 0.106 M for DifFeat, reflecting the deeper encoder–decoder structure of U-Net versus the shallower CNN in DifFeat. This makes DifFeat substantially lighter in terms of model complexity. The U-Net model required approximately 58 min (3,538 s) to complete 100 epochs using 528 labeled training images. In contrast, the DifFeat clustering model completed 100 epochs in just 4.61 s on a single unlabeled image tile, a 99.87% reduction in training time. Both models were trained using 512 × 512 × 6 (six-band composite) input slices to ensure consistency in spectral inputs for comparison. These results highlight the substantial contrast in model size and training efficiency between the two approaches.

Figure 3 presents qualitative comparisons between U-Net and DifFeat predictions across five representative Arctic coastal scenes. Each row (A–E) shows a WV2 scene, the manually annotated land (green), water (blue), and background (cream) reference mask, and the predicted segmentation results from both models using different spectral inputs. Columns three and four show U-Net predictions trained on RGB and RGB + IR + NDWI + NDVI inputs, respectively, while columns five and six display DifFeat predictions under the same input configurations. The labels “S” and “B” denote shoreline and bluff edge similarity scores, respectively, computed as the Intersection over Union (IoU) between each model’s predicted boundaries and the ground-truth reference (shown for visualization only).

Across the examples, U-Net produced smooth, continuous land and water regions, closely following the manually annotated masks in most cases. Adding infrared and normalized indices improved delineation of subtle or low-contrast transitions, particularly where shallow water or moist tundra obscure spectral differences. In contrast, DifFeat generated more granular clusters with greater local variability, yet effectively captured fine-scale transition zones between land and water even without labeled training data. Multispectral inputs improved both models, yielding better spatial coherence and reduced false detections. To illustrate, Rows C and E of Figure 3 demonstrate the influence of surface texture and illumination, where both models exhibited small gaps or over-segmentation in darker regions when using RGB inputs. The addition of infrared and normalized indices mitigated these artifacts, improving consistency and reducing misclassification.

Figure 4 presents the corresponding shoreline and bluff edge boundaries derived from the segmentation outputs shown in Figure 3. The first column shows the same WV2 input scenes for reference, while the subsequent panels display the extracted boundaries corresponding to each segmentation result in the same spatial arrangement. Blue lines represent the shoreline and orange lines represent the bluff edge, both automatically delineated from the predicted masks using the interface extraction algorithm described in Section 2.8. The visual comparison highlights that U-Net boundaries were generally smoother and more continuous, whereas DifFeat boundaries were finer and often better aligned with subtle geomorphic transitions, though occasionally fragmented. When infrared and normalized indices were included, both models achieved improved alignment with the ground-truth boundaries with reduced false positives. Together, these results demonstrate that both models can generate accurate land/water segmentation, and corresponding shoreline/bluff edge boundaries, with U-Net favoring spatial smoothness and DifFeat excelling in fine-scale texture capture, providing complementary strengths for Arctic coastal monitoring.