A new framework for selecting observation points and reconstructing wave fields under sparse observations

Abstract

In the context of increasingly frequent typhoons, tropical cyclones, and severe coastal storms that pose growing risks to maritime safety and offshore infrastructure, accurate reconstruction of ocean wave fields under sparse observation conditions has become a critical yet underexplored challenge. We propose a hybrid neural network model that integrates physical prior knowledge into a deep learning framework to optimize key observation point selection and enable high-accuracy reconstruction of wave statistics. The model comprises a U-Net-based decision network (Actor) for selecting observation points and a U-Net–GAN-based reconstruction network (Critic) for wave field recovery. A hybrid loss function incorporating physical constraints and region-specific sensitivity heatmaps guides the model toward high-impact observation areas, while spatial clustering strategies ensure broad spatial coverage. The closed-loop optimization mechanism leverages reconstruction error feedback to iteratively refine both observation strategies and reconstruction performance. Experiments using hourly multi-variable ERA5 reanalysis data in the South China Sea demonstrate that, under sparse observation settings, our approach significantly outperforms conventional deployment strategies in reconstruction accuracy, validating its effectiveness for resource-constrained marine monitoring applications.