Single-photon lidar observation process modeling and physics-driven virtual data generation

Single-photon lidar possesses significant advantages in underwater weak-signal target detection. However, the complex underwater light propagation environment leads to highly random and non-uniform observation noise. Furthermore, the limited scale of real data, constrained by experimental conditions and acquisition costs, makes it difficult to meet the training requirements of deep learning detection models. To address these issues, this paper proposes a physics-driven method for single-photon lidar observation modeling and high-fidelity data generation. First, a noise physics model is established based on the water photon detection mechanism to achieve parameterized characterization of the underwater observation degradation process. Then, virtual data matching the distribution of real observation data are synthesized through domain-adaptive physics optimization. Simultaneously, a simplified data synthesis method based on statistical feature constraints is proposed, enabling rapid construction of large-scale samples—generating over 14 000 images within 0.2 h. Experimental results demonstrate that the relative alignment index exceeds 90% across all evaluated scenarios, validating that both methods effectively balance data fidelity and generation efficiency to provide a robust data foundation for spectral power distribution target detection.