Coastline detection in satellite imagery: A deep learning approach on new benchmark data

Detailed and up-to-date coastline morphology data underpins our understanding of coastline change over time. The development of an automated and scalable coastline extraction methodology from satellite imagery is currently limited by the low availability of open, globally distributed and diverse labelled data with which to develop and benchmark techniques. Therefore, in this study we present the Sentinel-2 Water Edges Dataset (SWED), a new and bespoke labelled image dataset for the development and bench-marking of techniques for the automated extraction of coastline morphology data from Sentinel-2 images. Composed of 16 labelled training Sentinel-2 scenes, and 98 test label-image pairs, SWED is globally distributed and contains examples of many different coastline types and natural and anthropogenic coastline features. To provide a baseline of model performance against SWED we train and test four convolutional neural network models, based on the U-Net model architecture. Models are optimised using Categorical Cross-entropy Loss, Sørensen–Dice Loss and two novel loss functions we present for the focusing of model training attention to the boundary between land and water. Through a hybrid quantitative and qualitative model assessment process we demonstrate that the model trained using our novel Sobel-edge loss function has greater sensitivity to fine-scale, narrow coastline features whilst possessing near top quantitative performance demonstrated by Categorical Cross-entropy. The SWED dataset is published openly for use by the remote sensing and machine learning communities, whilst the Sobel-edge loss is available for use in machine learning applications where sensitivity to boundary features is important. • We describe a new dataset for training and benchmarking coastline extraction models. • The new dataset contains labelled Sentinel-2 imagery, in training and test splits. • The new dataset contains diverse coastline types and features from around the world. • Common and novel loss functions are used to optimise a convolutional neural network. • The novel Sobel-edge loss function showed greatest sensitivity to coastal features.