In this paper, we present a method for investigating the coupling of structure-borne sound (SBS) and underwater sound (UWS) of ships based on a decomposition in optimal frequency bands. The coupling is modeled by a linear regression model for sequences of spectral power, integrated over frequency intervals and moving time windows. As part of this analysis, a novel approach to detecting frequency intervals that bear particular relevance for the relationship between SBS and UWS is introduced. Regression models are regularized by a maximum-significance constraint. Cross correlations between UWS regressands and SBS regressors seem to be promising as quantitative measures for describing the coupling between SBS and UWS time series. The method is evaluated with experimental data recorded simultaneously from fixed SBS sensors onboard the research vessel R/V PLANET and from a freely drifting buoy system during an open-sea experiment north of the Outer Hebrides, Scotland, in 2011.