Physics-Guided Spatiotemporal Learning for Coastal Wave Peak Period Estimation from Video

Problem
The paper addresses the limitations of traditional coastal wave monitoring systems, which are often expensive and have restricted spatial coverage. Existing deep learning methods for passive ocean monitoring lack physical interpretability and validation in oceanographic contexts. This work proposes a novel approach to estimate nearshore wave peak periods directly from video, filling a gap in the literature regarding cost-effective and physically interpretable monitoring techniques. Notably, this is a preprint and has not undergone peer review.

Method
The authors introduce a Physics-Guided Deep Spatiotemporal Learning Framework that integrates several advanced techniques. The framework employs automated temporal-variance based region-of-interest detection to focus on relevant video segments. It utilizes multi-stage Sim-to-Real transfer learning to adapt models trained on synthetic data to real-world scenarios. The architecture includes both transformer-based and recurrent-convolutional networks, with a focus on lightweight designs for operational efficiency. Physics-informed regularization is applied to enhance predictive accuracy and ensure physical consistency in the model outputs. The training process involves synthetic pretraining, silver-label adaptation, and expert fine-tuning to optimize performance.

Results
The proposed framework demonstrates superior performance in estimating wave peak periods compared to baseline methods. Specifically, transformer-based architectures achieved higher accuracy in instantaneous predictions, while lightweight recurrent-convolutional architectures exhibited enhanced temporal stability and operational oceanographic skill. The ablation studies indicate that physics-guided regularization significantly improves trend-following consistency and reduces the occurrence of physically implausible predictions. The explainability auditing process confirmed that the model effectively identifies hydrodynamically active surf-zone regions, aligning well with established wave propagation behaviors.

Limitations
The authors acknowledge that the framework’s reliance on video data may limit its applicability in conditions where video capture is not feasible. Additionally, while the physics-informed regularization improves model reliability, the complexity of the model may still lead to challenges in interpretability for end-users. The study also does not address the potential impact of varying environmental conditions on model performance, which could affect generalizability.

Why it matters
This research has significant implications for coastal engineering and management, offering a cost-effective alternative to traditional monitoring systems. The integration of physics-informed deep learning with video data could enhance long-term coastal wave monitoring, contributing to improved shoreline protection and marine hazard assessment. The findings support the development of more accessible and interpretable oceanographic tools, as discussed in related works on deep learning applications in environmental monitoring, as published in arXiv.