Reconstruction of sparse ocean noise fields with generative neural networks

Modeling the underwater noise field offers important insights for underwater navigation, assessing noise impacts, and establishing noise budgets. In situ measurements of the noise field are often sparse with limited information about spatially distributed underwater acoustic fields. In this work, we demonstrate the use of a generative neural network for estimating noise fields over regions not directly measured. An ARiA developed ambient noise model is used to generate noise fields for various ocean noise-source scenarios. The model output includes vertical directionality, depth dependence, and incorporates multiple sources of sound, including anthropogenic, biologic, and environmental. We train generative models on these simulated spatial noise distributions a priori. These generative models are able to take sparsely sampled noise measurements and construct the spatial distribution of the noise field. The generative model learns the spatial distribution by minimizing the point-to-point difference between the generated fields and the true fields at each step while also maximizing the fidelity of the general distribution of the generated noise fields to the true fields. Through a nested generative model approach, along with auxiliary sound-speed-profile information, the local field is used to extrapolate the noise field to larger spatial scales.