On the Role of Data Augmentation in Supervised Contrastive Learning for Coastal Erosion Prediction

Accurate prediction of shoreline change is crucial for informed coastal management, hazard mitigation, and climate resilience. This study presents a deep learning framework that leverages Supervised Contrastive Learning (SupCon) to extract robust representations from satellite coastal imagery. We employ MobileNet, a family of lightweight yet powerful convolutional neural networks, as the backbone to support SupCon's discriminative feature learning. A central focus of our investigation is the role of data augmentation, a key driver in contrastive learning performance, evaluated through a comprehensive ablation study. The proposed approach significantly improved prediction accuracy at 81.49% for MixUp and 83.77% for CutMix, compared to the traditional supervised deep learning approach at 61.69% and basic supervised contrastive learning at 64.29%. This development highlights the effectiveness of data augmentation techniques for improving the performance of SupCon for downstream classification tasks.