EnCo SupCon: Entropy-driven supervised contrastive learning for discriminative feature representations from remote sensing imagery

• An entropy-based coordinate supervised contrastive learning approach is proposed to extract enhanced feature representations from remote sensing imagery. • ResNet50V2 achieved the highest accuracy of 99.74% with manual augmentations, reflecting the benefits of deeper architectures for capturing complex spatial and spectral shoreline patterns. • For computationally efficient deployment, MobileNetV2 combined with advanced augmentation strategies such as CutMix achieved 98.46% accuracy, approaching the performance of deeper networks while significantly reducing training time and resource requirements. Coastal erosion poses serious challenges, including damage to infrastructure, loss of farmland, and ecosystem degradation. Monitoring and predicting coastal changes are critical for mitigation, but existing machine learning approaches often struggle to generalize across heterogeneous coastal environments and require extensive labeled data for reliable performance. To address these challenges, we propose Entropy-based Coordinate Supervised Contrastive Learning (EnCo SupCon), a deep learning framework designed to extract robust, spatially-aware feature representations from satellite imagery. The model integrates three key components: (i) entropy weighting to emphasize informative, near-boundary regions, (ii) coordinate attention to capture horizontal and vertical spatial dependencies, and (iii) supervised contrastive learning to cluster similar features while separating dissimilar ones in the embedding space. Ablation studies demonstrate that entropy weighting and coordinate attention individually improve performance and act synergistically when combined, raising MobileNetV2 accuracy from 77.79% to 90.78%. Data augmentation further enhances performance, with CutMix achieving 98.46%, outperforming MixUp (83.51%) and manual augmentations (90.78%), by encouraging the model to attend to multiple discriminative regions simultaneously. Sensitivity analysis shows that moderate entropy weights and lower temperature values provide stable and high accuracy, while Grad-CAM visualizations confirm that CutMix combined with entropy-guided coordinate attention focuses on the most informative coastal features. EnCo SupCon generalizes effectively to geographically distinct coastal datasets, with high accuracy across deeper backbones, while lightweight architectures maintain competitive performance with lower computational cost. Statistical analysis indicates that augmentation strategies provide the greatest benefit for smaller models, highlighting the interaction between architecture, entropy weighting, and augmentation for robust coastal monitoring.