PaperlayerView exact revision on arXiv

Source-linked AI summary

HED-UNet: Combined Segmentation and Edge Detection for Monitoring the Antarctic Coastline

Konrad Heidler, Lichao Mou, Celia Baumhoer, Andreas Dietz, Xiao Xiang Zhu

arXiv:2103.01849v1cs.CVeess.IV

TL;DR

Antarctic coastline detection requires both sea-land segmentation and precise edge delineation, yet existing models commonly treat them separately. HED-UNet unifies the tasks with multiresolution prediction, deep supervision, and hierarchical attention, and it outperforms the evaluated baselines while adding little computational cost. The method is intended specifically for coastline detection rather than as a general-purpose model.

  • Problem

    Antarctic coastline monitoring is difficult because dynamic fronts and imagery challenges complicate delineation, while existing approaches often focus on either sea-land segmentation or coastline detection.

  • Method

    HED-UNet jointly predicts segmentation masks and coastline edges using a UNet- and HED-inspired architecture with deep supervision and hierarchical attention merging.

  • Results

    HED-UNet surpasses edge-detection and semantic-segmentation baselines, with especially strong average-deviation performance in the Antarctic Peninsula validation area.

  • Takeaways & Limitations

    Combining segmentation and edge detection helps HED-UNet better grasp the coastline concept, with little additional computational cost compared with regular UNet.

  • Takeaways & Limitations

    The proposed model is not a general-purpose model, and its validation results still contain imperfect cases with two reported failure modes.

Abstract

from arXiv · show

Deep learning-based coastline detection algorithms have begun to outshine traditional statistical methods in recent years. However, they are usually trained only as single-purpose models to either segment land and water or delineate the coastline. In contrast to this, a human annotator will usually keep a mental map of both segmentation and delineation when performing manual coastline detection. To take into account this task duality, we therefore devise a new model to unite these two approaches in a deep learning model. By taking inspiration from the main building blocks of a semantic segmentation framework (UNet) and an edge detection framework (HED), both tasks are combined in a natural way. Training is made efficient by employing deep supervision on side predictions at multiple resolutions. Finally, a hierarchical attention mechanism is introduced to adaptively merge these multiscale predictions into the final model output. The advantages of this approach over other traditional and deep learning-based methods for coastline detection are demonstrated on a dataset of Sentinel-1 imagery covering parts of the Antarctic coast, where coastline detection is notoriously difficult. An implementation of our method is available at \url{https://github.com/khdlr/HED-UNet}.

I. INTRODUCTION

Antarctica’s dynamic, difficult-to-observe coastline makes automated monitoring necessary, while coastline delineation intrinsically combines sea-land segmentation and edge detection. HED-UNet addresses this duality with multiresolution predictions, deep supervision, and hierarchical attention.

  • Antarctica’s coastline changes through iceberg calving and seasonal and climate variations, making continuous monitoring important for glaciological processes and sea-level contribution.
  • Around 40 000 km of Antarctic coastline makes manual delineation infeasible, especially across multiple time steps.
  • SAR coastline extraction is challenged by speckle and seasonally varying glacial-ice backscatter, requiring contextual rather than purely local information.
  • Human delineation jointly considers sea-land areas and their separating edge, motivating a model trained simultaneously for segmentation and edge detection.
  • HED-UNet predicts segmentation masks and coastline edges at multiple resolutions using an encoder-decoder architecture with skip connections.
  • Deep supervision improves training efficiency and generalization, while hierarchical attention blends coarse and high-resolution predictions according to information across levels.

II. RELATED WORK

Antarctic coastline detection is complicated by dynamic sea ice and visually similar features that require spatial context. Existing approaches include statistical, unsupervised, segmentation, and coastline-focused methods with distinct strengths and limitations.

  • Antarctic coastline methods must account for icebergs and ice mélange, which can resemble land ice locally and require spatial context for exclusion.
  • Coastline-detection approaches generally divide into methods predicting sea-land areas and methods predicting the coastline itself.
  • Statistical thresholding is fast and easy to implement but discards spatial relationships, limiting its ability to handle complex imagery issues.
  • Unsupervised methods require no prior training, but their lack of supervision prevents teaching them to ignore features such as icebergs.
  • Deep convolutional models use encoder-decoder context and can outperform pixelwise or shallow texture-based methods, but generally require substantial labeled data and training time.

1) Edge Tracing:

Coastline-related vision methods range from hand-designed edge operators to learned and hybrid segmentation approaches. Their limitations motivate combining precise edge information with contextual area predictions.

  • 1) Edge Tracing:: Classical edge tracing filters likely edge pixels and connects them into boundaries, including through shortest-path algorithms or ridge tracing.
  • 1) Edge Tracing:: Wrong turns during tracing are difficult to correct, creating robustness problems for these relatively simple methods.
  • 1) Edge Tracing:: Active contours iteratively deform an initialized curve, but sensitive initialization can lead to local minima that do not represent the desired edge.
  • 1) Edge Tracing:: Deep CNNs can outperform handcrafted Roberts and Sobel operators by learning edge features through deeper networks.
  • 1) Edge Tracing:: Semantic segmentation often becomes blurry near class boundaries because boundary pixels are a minority under standard pixelwise cross-entropy loss.
  • 1) Edge Tracing:: Auxiliary edge losses, edge outputs, propagation, or edge-derived features have been used to sharpen or improve segmentation results.
  • 1) Edge Tracing:: The proposed unified theory treats segmentation and edge detection as equally important components of coastline detection.

III. PROPOSED METHOD

HED-UNet is a hybrid network for simultaneous sea-land segmentation and coastline-edge prediction, combining contextual encoder-decoder processing with multiresolution task heads. Its design addresses the blur and ambiguity of using either task alone.

  • III. PROPOSED METHOD: UNet-based sea-land segmentation performs well broadly but can become inaccurate and blurry near the coastline, where precise location is central.
  • III. PROPOSED METHOD: Edge detection provides sharp boundaries but cannot label inside versus outside and can be confused by inland structures or nearby icebergs.
  • III. PROPOSED METHOD: HED-UNet jointly predicts sea-land segmentation and coastline edges, combining the complementary advantages of both tasks.
  • III. PROPOSED METHOD: The high-level framework computes a pyramid of feature maps, then task-specific merging heads combine them using hierarchical attention.
  • III. PROPOSED METHOD: Its backbone aggregates contextual information through an encoder-decoder structure, while prediction heads support multiple task outputs.

2) Edge Detection Building Blocks:

The model combines segmentation and edge-detection components through an encoder-decoder feature pyramid with multiresolution side outputs and task-specific merging heads.

  • 2) Edge Detection Building Blocks:: Edge detection frameworks combine sharp high-resolution edges with robust lower-resolution features to suppress false positives.Their encoders aggregate contextual information by downsampling, while merging blocks combine upsampled representations across resolutions.
  • 2) Edge Detection Building Blocks:: The proposed high-level architecture contains components for both semantic segmentation and edge detection.The network structure is derived from the complementary building blocks of the two task families.
  • 2) Edge Detection Building Blocks:: An adaptively balanced binary cross-entropy loss gives equal weight to positive and negative classes for both tasks.This addresses the strong edge-versus-no-edge class imbalance and is also used for semantic segmentation.
  • 2) Edge Detection Building Blocks:: The feature pyramid uses 6 resolution levels, spanning full image resolution to 1/32 resolution through 5 down- and upsampling steps.The six levels provide the large spatial context window used for coastline decisions.
  • 2) Edge Detection Building Blocks:: Each feature-map level produces side outputs for both segmentation and edge detection, yielding 6 multiscale outputs.These outputs support both deep supervision and subsequent merging.

1) Deep Supervision:

Deep supervision trains intermediate multiresolution predictions, while hierarchical attention merges them into a full-resolution output using input-dependent resolution weights.

  • 1) Deep Supervision:: Deep supervision trains each intermediate prediction against an appropriately downsampled ground truth.Additional loss terms provide more direct feedback to earlier layers and are intended to improve learning effectiveness and generalization.
  • 1) Deep Supervision:: The side outputs become inputs to merging heads that combine fine-grained edge-region predictions with robust higher-level predictions away from edges.This multiscale combination follows the central merging idea of HED.
  • 1) Deep Supervision:: Hierarchical attention dynamically fuses multiscale predictions according to the confidence associated with different granularities.Unlike fixed fusion weights, the mechanism lets the network attend to resolution levels useful for each pixel.
  • 1) Deep Supervision:: For each prediction level, weight maps are upsampled, normalized across levels with softmax, and combined with predictions through a dot product.The resulting operation produces the final full-resolution prediction.
  • 1) Deep Supervision:: The attention merger is interpreted as a mixture model whose coefficients depend on the input rather than remaining fixed.This input dependence provides a more flexible probabilistic model with potential for better classifications.

IV. DATASET AND EXPERIMENTAL SETUP

The study evaluates HED-UNet and competing methods on manually annotated Sentinel-1 Antarctic coastline scenes, using geographically separated training and validation data.

  • IV. DATASET AND EXPERIMENTAL SETUP: The experiments train and validate several competing methods alongside the proposed model on the Antarctic coastline dataset.This setup is used to validate the suggested model improvements.
  • IV. DATASET AND EXPERIMENTAL SETUP: The dataset contains 16 cropped Sentinel-1 GRD scenes acquired from June 2017 through December 2018 in Extra Wide Swath mode.The imagery uses 40 m resolution and dual HH/HV polarization channels.
  • IV. DATASET AND EXPERIMENTAL SETUP: Experts manually annotated the scenes to provide ground truth for sea-land segmentation and coastline delineation.The scenes cover approximately 730 000 km2 with average cropped dimensions of 7870 × 6572 pixels.
  • IV. DATASET AND EXPERIMENTAL SETUP: The geographic split uses 11 training scenes and 5 validation scenes, with two areas completely excluded from training.Tiles are 768 × 768 pixels with 50% overlap, and training uses 8-fold data augmentation.

B. Evaluated Models

The evaluation compares traditional, segmentation, edge-detection, and modern deep-learning models, using qualitative validation imagery and numerical results.

  • B. Evaluated Models: The baseline set includes Gaussian Mixture, K-Medians Clustering, Sobel Edges, and Active Contours methods.These provide traditional unsupervised, filtering, and contour-based comparisons.
  • B. Evaluated Models: Deep-learning competitors include HED, UNet, DeepUNet, RDUNet, HRNet + OCR, and Gated-SCNN.They represent edge detection, sea-land segmentation, and state-of-the-art semantic segmentation approaches.
  • B. Evaluated Models: The deep-learning models are trained for 15 epochs with Adam on an Nvidia V100 GPU using batches of 4 samples.The optimizer uses learning rate 0.001, β1 = 0.9, β2 = 0.999, and ε = 10^-8.
  • B. Evaluated Models: Qualitative comparisons use unseen validation tiles selected across the full range from easy scenes at the top to hard scenes at the bottom.This visualization is intended to represent scene difficulty variation in the validation set.
  • B. Evaluated Models: Table I reports numerical results for the evaluated models.The table is the quantitative counterpart to the model comparison setup.

V. RESULTS AND DISCUSSION

HED-UNet outperforms alternative approaches on average coastline deviation, while model performance varies substantially between validation regions. Ablations show that deep supervision is crucial for edge detection, whereas the full model’s improvements are task-specific.

  • Model comparison: Traditional models are not competitive on this dataset, which the authors attribute to icebergs and difficult ice-sheet backscatter characteristics.Because these models are unsupervised, they cannot learn how to handle such impediments.
  • Model comparison: All evaluated models perform worse on the Antarctic Peninsula than in Wilkes Land, indicating strong regional variation in difficulty.The Antarctic Peninsula is described as a particularly tough validation location.
  • Model comparison: UNet provides a respectable deep-learning baseline and outperforms HRNet+OCR and Gated-SCNN on some evaluated metrics.Gated-SCNN’s edge results fall short of HED and HED-UNet in the reported comparison.
  • Model comparison: HED-UNet achieves the best average deviation among evaluated approaches, particularly in the Antarctic Peninsula validation area.The study identifies average distance as the most important metric because it estimates overall coastline error.
  • Ablation studies: Reducing the network from six to five resolution levels performs slightly worse than full HED-UNet but still outperforms baseline methods.This comparison uses five resolution levels to match UNet-based models.
  • Ablation studies: Removing deep supervision severely harms edge detection and can cause the coastline to be missed, while semantic segmentation changes little.The authors describe deep supervision as of paramount importance for edge detection performance.

C. Merging Strategies

The model compares no merging, learned merging, and hierarchical attention merging for multiresolution predictions. Attention dynamically adapts fusion across scenes and spatial regions, while DEM input can help difficult scenes but may cause overfitting.

  • Merging strategies: The ablation compares no merging, learned merging, and attention-based merging strategies for combining predictions across resolution levels.The no-merging configuration uses only the final decoder layer, while learned merging combines upsampled predictions with a 1×1 convolution.
  • Merging strategies: Learned merging provides little segmentation improvement, worsens edge detection slightly, and changes average deviation differently across the two validation regions.The authors suggest fixed coefficients may suit Wilkes Land better than the Antarctic Peninsula.
  • Merging strategies: Attention merging computes pixelwise weights dynamically, allowing different merging coefficients for Wilkes Land and the Antarctic Peninsula.This adaptability is presented as overcoming the scene-specific weakness of learned merging.
  • Hierarchical attention: The attention mechanism uses lower-resolution predictions more on homogeneous regions and emphasizes the highest resolution near edges.The reported behavior combines robust coarse predictions with accurate fine delineations.
  • DEM experiments: DEM inclusion can cause overfitting when elevation and SAR imagery contradict each other in dynamic coastal regions.The authors therefore state that DEM data must be used carefully to prevent reliance on the DEM alone.

E. Limitations

The model’s remaining errors mainly arise from large sea-ice clusters and insufficient context near tile borders. Effective receptive-field analysis suggests that larger spatial context is associated with better validation scores, while failures persist in especially challenging scenes.

  • 1) Sea Ice:: Large clusters of sea ice can still confuse HED-UNet despite its large receptive field and multitask training.The model’s first failure mode involves sea ice that is misclassified because it resembles land ice.
  • 2) Missing Context:: Tile-border regions are sometimes misclassified because the model lacks enough contextual information.One example shows sea ice next to a tile border wrongly classified as land.
  • E. Limitations: The reported failure modes are infrequent and also affect the other compared models.The authors note that large-context interpretation is particularly difficult for neural networks without general reasoning capabilities.
  • F. Effective Receptive Fields: Effective receptive fields measure which input pixels influence a central output prediction through expected gradient magnitude.The spatial distribution of influential pixels represents the context actually used by the CNN, rather than its theoretical receptive field alone.
  • F. Effective Receptive Fields: Larger effective receptive fields appear correlated with better validation scores for Antarctic coastline detection.HED-UNet and HRNet+OCR employ very large effective receptive fields, supporting the need for broad spatial context.

VI. CONCLUSION

HED-UNet jointly performs segmentation and edge detection, using deep supervision and attention merging to combine multiscale predictions. It surpasses segmentation and edge-detection baselines for coastline detection while adding little computational cost, though it is not a general-purpose model.

  • VI. CONCLUSION: HED-UNet simultaneously performs semantic segmentation and edge detection, exploiting synergies between the two tasks.The model is designed specifically to combine coastline-region prediction with precise edge delineation.
  • VI. CONCLUSION: Deep supervision and attention merging encourage meaningful deep features and more robust classifications.The architecture uses additional training feedback and attention-based merging heads to improve learning performance.
  • VI. CONCLUSION: HED-UNet surpasses both edge-detection and semantic-segmentation baselines for Antarctic coastline detection.The conclusion attributes the improvement to combining the two tasks in a unified model.
  • VI. CONCLUSION: The architecture requires little additional computational cost compared with a regular UNet.Most performance gains are attributed to the adapted training procedure and a few additional layers.
  • VI. CONCLUSION: HED-UNet is beneficial for coastline detection but is not presented as a general-purpose model.The authors suggest possible broader applications, including building footprints, roads, lakes, and rivers.
Loading 2103.01849v1…