Physics-informed waveform decomposition with scene adaptation for airborne laser bathymetry

Airborne LiDAR bathymetry (ALB) is essential for coastal mapping. However, severe signal degradation in complex optical environments often causes unstable depth retrieval. To address this, we propose a physics-informed waveform decomposition (PI-DC) framework. PI-DC integrates a differentiable, radiative-transfer-guided forward model directly into the backpropagation process, thereby addressing the limitations of unconstrained neural networks. The framework features a scene-aware neural initialization (SANI) module that enforces physical consistency by employing temporal decoupling to resolve overlapping ultra-shallow echoes and dynamic signal gating to suppress spurious seabed detections in deep waters. These predictions establish reliable physical boundaries for a scene-adaptive box-constrained refinement (SABR) module, ensuring robust parameter convergence via a Levenberg–Marquardt (LM) optimizer. Evaluations on simulated and real UAV-borne datasets demonstrate that PI-DC consistently outperforms traditional model-based and purely data-driven approaches. Specifically, it reduces the overall depth RMSE to 0.045 m in simulations while yielding more stable and physically plausible results on real-world UAV-borne measurements.