Answers are grounded in this paper and cite their sources.
Satellite-derived bathymetry faces two coupled obstacles: labels are scarce, and models transfer poorly across regions. We propose RT-ProtoNet, a radiative-transfer-guided few-shot prototype network, which takes patches from three penetrating Sentinel-2 bands and treats ICESat-2 depths, aggregated to the 10 m grid, as sparse supervision; cross-region episodic meta-training then yields a transferable metric representation. Rather than embedding the radiative transfer relation as an analytic formula, we use it as a comparability criterion between support and query pixels, entering the attention as a learnable additive distance penalty. The prediction sums a convex combination over support depths and a base regression head trained on source domains only and frozen at deployment, so no target-domain parameters are estimated. We evaluate on five optically clear shallow reef regions by leave-one-region-out cross-validation against airborne laser bathymetry, against nine baselines, where K is the total number of support pixels (3, 5, 10, 20, 50). For K ≤ 20, RT-ProtoNet attains the lowest cross-region macro-mean RMSE (2.79–2.48 m), with a run-to-run standard deviation 0.67 times the runner-up’s; at K = 50, pre-training with head fine-tuning overtakes it. Ablation shows the residual head’s gain scales with support depth coverage (correlation 0.991) and that the two physical priors are largely substitutable; twelve perturbations to the attenuation input change RMSE by only 0.66%. The method suits depths of roughly 2–20 m and budgets of K ≤ 20, and does not extrapolate to turbid or optically complex waters. Independent evaluation at two external optical-domain sites, a moderately turbid embayment (Tampa Bay) and a clear-to-moderately-turbid transition site (Boquerón Bay), further delineates the method’s applicability across contrasting optical conditions, showing reliable transfer within the clear-water regime but substantial degradation under stronger optical attenuation.
No takes yet. Share an insight, caveat, or question.
Kong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: