Inversion of multimegameter-range acoustic data for ocean temperature

Abstract

Ocean sound speed (a surrogate for temperature) derived from the ray travel times obtained from acoustic transmissions may be inaccurate when the reference ocean state is inadequate for linearized inversion. When the reference (e.g., the Levitus ocean atlas) is significantly different from the "true" ocean, the reference ray paths inaccurately represent the true sampling. In addition, natural oceanic variations, such as the evolution of a summer mixed layer, can significantly change the ray sampling over time. The guiding principle for inversion is that ray travel times associated with the inverse solution must match the measured travel times. A time-dependent reference ocean can reduce both the nonlinearities and the solution uncertainties since the model variances may be assumed to be less. The Levitus ocean atlas was employed to explore the effects of nonlinearities when inverting multimegameter-range acoustic data and to find accurate inversion methods. These methods were applied to acoustic data obtained in the North Pacific during the acoustic thermometry of ocean climate (ATOC) project using an acoustic source on Pioneer Seamount off the coast of California. In order to linearize the inversions, the annual cycle was removed by referencing the measured travel times to travel times computed using the Levitus monthly ocean atlas. This linearization results in a more accurate time series of range- and depth-averaged temperatures, but the solution for range- and depth-averaged temperature is only slightly different from that using a time-independent set of rays. Standard uncertainties for the 0-1000-m depth-averaged temperature are typically /spl plusmn/0.012/spl deg/C, while the annual peak to-peak temperature variation is about 0.4/spl deg/C. Because the travel time data are inherently averaging, the time series of range- and depth averaged temperature is insensitive to different assumptions made in the forward model, such as the model parameterization, variances, wavenumber spectra, and the data uncertainties.