Deep CNNs beat Random Forest for scalable satellite-derived ocean depth mapping

A new paper from researchers at Ohio State University evaluates machine learning and deep learning for transferable satellite-derived bathymetry (SDB) using Sentinel-2 multispectral imagery. The study compares a Random Forest baseline against four convolutional neural networks—ResNet-50, ResNet-101, EfficientNet-B4, and ConvNeXt-Large—trained on Pratas Island and selected Great Barrier Reef regions, then tested on spatially independent intra- and cross-regional areas. The single most impactful design choice was preserving spatial continuity by keeping contiguous reef blocks during training rather than random patches. Additionally, the authors introduced a Smooth Weight Function (SWF)-weighted RMSE loss that emphasizes near-surface depths, significantly improving shallow-water accuracy.

Results show intra-regional RMSE ranging from 1.15 to 1.92 m over the 0–20 m depth range, dropping to just 0.26 m for depths ≤3 m. Under cross-regional transfer, Random Forest degraded sharply (RMSE from 1.53 m to 2.99–3.78 m), while deep models stayed more robust (2.46–2.98 m). On the public MagicBathyNet aerial-RGB benchmark (0–16 m), the proposed CNNs achieved 0.19–0.22 m RMSE, outperforming both a U-Net baseline and a task-specific transformer architecture with substantially fewer parameters. The study also exploited multi-temporal repeat imagery, finding that training on multiple passes broadens diversity and that median-aggregating predictions reduces noise from changing sun angles, atmospheric conditions, water properties, and tides.

The authors have released optimized architectures and pretrained weights to enable scalable transfer to new coastal sites. This work demonstrates that deep learning, when carefully designed with spatial continuity and depth-aware losses, can deliver accurate bathymetry maps across diverse regions without local retraining—a significant step toward global coastal monitoring.