From Local Training to Large-Scale Mapping: A Comparative Assessment of Machine Learning and Deep Learning for Transferable Satellite-Derived Bathymetry

Satellite-derived bathymetry (SDB) provides a cost-effective means for mapping shallow-water depths, yet its scalability and cross-regional generalizability remain challenging in optically complex coastal environments. This study systematically evaluates machine learning (ML) and deep learning (DL) approaches for transferable SDB over the 0–20 m depth range using multispectral Sentinel-2 imagery. A Random Forest model and four deep learning architectures–ResNet-50, ResNet-101, EfficientNet-B4, and ConvNeXt-Large–are developed and trained using data from Pratas Island (South China Sea) and selected reef regions of the Great Barrier Reef (GBR), and subsequently evaluated on spatially independent intra-regional and cross-regional test areas to assess generalization performance. Model sensitivity is investigated with respect to key training configurations, including loss-function design and data-splitting strategy. To enhance shallow-water learning, we introduce a Smooth Weight Function (SWF)-weighted RMSE loss that emphasizes near-surface depths and compare it with conventional RMSE and relative percentage error (RPE) objectives. In terms of training data, preserving spatial continuity during training substantially improves both numerical accuracy and structural consistency of predictions compared with random patch splitting. While the Random Forest model performs competitively in intra-regional tests, its accuracy degrades under cross-regional transfer (RMSE increasing from 1.53 m to 2.99–3.78 m). Deep learning models, although not always outperforming Random Forest in intra-regional settings, exhibit greater robustness to geographic shift. Using the spatially continuous training strategy, intra-regional RMSE ranges from 1.15 to 1.92 m over the full 0–20 m range, with shallow-water RMSE as low as 0.26 m for depths ≤ 3 m. Cross-regional transfer to geographically independent reefs yields moderate RMSE values of approximately 2.46–2.98 m (0–20 m range), indicating that geographic transfer remains challenging despite meaningful improvements over Random Forest. We further benchmark the proposed architectures against a task-specific bathymetry network using the public MagicBathyNet dataset. Under a unified 0–16 m shallow-water configuration using aerial RGB imagery, the proposed models achieve RMSE values between 0.19 and 0.22 m, outperforming both the baseline U-Net and the transformer-based bathymetry architecture while using substantially fewer parameters. In addition, we exploit multi-temporal repeat imagery for both training and inference, which increases training diversity and improves robustness to temporal variability arising from changing sun angles, atmospheric conditions, water properties, and tides. During inference, predictions from multiple repeat images are aggregated using the median to reduce noise and improve stability. Finally, we release optimized network architectures and pretrained weights to facilitate scalable application to new sites. This work demonstrates a practical pathway toward transferable, large-area SDB from multispectral satellite imagery using deep learning.