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

TL;DR

A new model is devised to unite these two approaches in a deep learning model, 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.

Abstract

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}.

Figures (9)

• Figure 1: In coastline detection, the vision tasks of segmentation and edge detection are inseparable.
• Figure 2: High-level structure of the proposed framework. First, the encoder and decoder calculate a pyramid of feature maps. Then, the task-specific merging heads combine this information using the hierarchical attention mechanism.
• Figure 3: Architectural details of the proposed network. The full model contains two task-specific merging heads, for clarity, only the segmentation head is shown here. The edge detection head follows the same structure.
• Figure 4: Spatial distribution of the scenes in the dataset. Scenes marked in green were used for model training, scenes marked in red were used for validation purposes. The red area in the top left is the "Antarctic Peninsula" validation site, while the bottom right red area is the "Wilkes Land" validation site. For most locations, data from 2 or 3 different sensing dates was used to allow for an assessment of each model's temporal stability. Marked in yellow is the footprint of the visualization tile in Fig. \ref{['fig:largevis']}.
• Figure 5: Qualitative results comparing the evaluated models on unseen validation tiles. In order to provide an informative visualization, the visualized tiles were selected to represent the full spectrum of easy (top) to hard (bottom) scenes within the validation set.