Deep learning for low altitude coastline segmentation

Coastline segmentation is the process of separating the coastal and backshore zones on aerial images. With a large world population living close to the coast, monitoring coastline changes is critical. Classical computer vision techniques were used to segment the coastline in high quality grayscale images where the difference between the zones was easy to distinguish. However, these techniques are limited to low resolution images and in areas with similar colors or textures. In this work we propose deep convolutional architectures for coastline segmentation using aerial images. An F1 score above 96% was obtained by the best performing model. The obtained results show that our deep models are capable of automatically and accurately detecting coastlines which will help in speeding-up the coastline localization process in large aerial images and improve the efficiency of monitoring coastal areas.