The authors describe the application of an adaptive optimal kernel (AOK) time-frequency representation to the processing of underwater acoustic data. The optimal kernel is a signal-dependent radially Gaussian function. Examples are given which demonstrate the effectiveness of the approach for simulated and real sonar data. The simulations indicate that the technique should work well for a larger set of signal classes than any current fixed-kernel representation. The technique has excellent performance even in the presence of substantial additive noise; this property may be exploited for signal detection. The AOK technique appears to offer unique features that can be used to characterize and automatically classify signals of interest, particularly when compared to other processing techniques.