A meta-deep-learning framework for spatio-temporal underwater SSP inversion

Sound speed distribution, represented by a sound speed profile (SSP), is of great significance because the nonuniform distribution of sound speed will cause signal propagation path bending with Snell effect, which brings difficulties in precise underwater localization such as emergency rescue. Compared with conventional SSP measurement methods via the conductivity-temperature-depth (CTD) or sound-velocity profiler (SVP), SSP inversion methods leveraging measured sound field information have better real-time performance, such as matched field process (MFP), compressed sensing (CS) and artificial neural networks (ANN). Due to the difficulty in measuring empirical SSP data, these methods face with over-fitting problem in few-shot learning that decreases the inversion accuracy. To rapidly obtain accurate SSP, we propose a task-driven meta-deep-learning (TDML) framework for spatio-temporal SSP inversion. The common features of SSPs are learned through multiple base learners to accelerate the convergence of the model on new tasks, and the model’s sensitivity to the change of sound field data is enhanced via meta training, so as to weaken the over-fitting effect and improve the inversion accuracy. Experiment results show that fast and accurate SSP inversion can be achieved by the proposed TDML method