Semantic segmentation of coastal aerial/satellite images using deep learning techniques: An application to coastline detection

Abstract

A new CNN based approach supported by semantic segmentation, was proposed. This approach is frequently used to carry out regional-scale studies. The core of our method revolves around a CNN model, based on the famous U-Net architecture. Its purpose is to identify different classes of pixels on satellite images and later to automatically detect the coastline. The recently launched Coast Train dataset was used to train the CNN model. Traditional coastline detection was improved (“water/land” segmentation) by means of two new aspects the use of the Sobel-edge loss function and the segmentation of the satellite images into several categories like built-up areas, vegetation and land besides beach/sand and water classes. The approach used ensures a more precise coastline extraction, distinguishing water pixels from all other categories. Our model adeptly identifies features, such as cliff vegetation or coastal roads, that some models might overlook. In this way, coastline localization and its drawing for regional scale study, have minor uncertainties. The performance of the CNN-based method, achieving 85% accuracy and 80% IoU (Intersection over Union) in the segmentation process. The ability of the model to extract the coastline was validated on a Sicilian case study, notably the San Leone beach (Agrigento). The model's results align closely with the ground truth, moreover, its reliability was further confirmed when it was tested on other Sicilian coastal regions.

Beyond robustness, the model offers a promising avenue for enhanced coastal analysis potentially applicable to coastal planning and management.