Development of a Shoreline Detection Method Using an Artificial Neural Network Based on Satellite SAR Imagery

Monitoring shoreline change is one of the essential tasks for sustainable coastal zone management. Due to its wide coverage and relatively high spatiotemporal monitoring resolutions, satellite imagery based on synthetic aperture radar (SAR) is considered a promising data source for shoreline monitoring. In this study, we developed a robust shoreline detection method based on satellite SAR imagery using an artificial neural network (NN). The method uses the feedforward NN to classify the pixels of SAR imagery into two categories, land and sea. The shoreline location is then determined as a boundary of these two groups of classified pixels. To enhance the performance of the present NN for land–sea classification, we introduced two different approaches in the settings of the input layer that account not only for the local characteristics of pixels but also for the spatial pixel patterns with a certain distance from the target pixel. Two different approaches were tested against SAR images, which were not used for model training, and the results showed classification accuracies higher than 95% in most SAR images. The extracted shorelines were compared with those obtained from eye detection. We found that the root mean square errors of the shoreline position were generally less than around 15 m. The developed method was further applied to two long coasts. The relatively high accuracy and low computational cost support the advantages of the present method for shoreline detection and monitoring. It should also be highlighted that the present method is calibration-free, and has robust applicability to the shoreline with arbitrary angles and profiles.