Coastline segmentation on Landsat 8 OLI images using majority voting with deep learning models

Abstract

Coastlines are highly dynamic due to both natural processes and anthropogenic factors, including global warming and sea level rise. Accurate coastline segmentation is essential for effective monitoring and management. Although previous studies have applied deep learning for coastline detection, many existing models still suffer from instability across scenes, blurred boundaries, and segmentation artifacts, indicating that model generalization remains a challenge. This study aims to develop a more robust coastline segmentation approach by introducing an automated majority voting strategy that integrates three deep learning models: ResNet50, ResNet18, and MobileNet-V2. Landsat 8 OLI imagery is used for training and testing. The Jaccard index results show that ResNet18, ResNet50, and MobileNet-V2 achieved scores of 0.96, 0.98, and 0.95 respectively, while the proposed majority voting method also achieved 0.98. Despite the producing a similar numerical score to the best individual model (ResNet50), the ensemble method improves segmentation consistency by reducing artifacts such as unwanted peripheral shapes and cracks within land areas. These findings demonstrate that combining multiple segmentation outputs yields more stable and reliable coastline detection than using single models. Future work will apply this approach to broader Indonesian coastal regions to further assess its generalizability across diverse shoreline conditions.