Frequency-Domain Second-Order Decorrelation with Compact Time-Domain Regularization for Convolutive Underwater Acoustic Source Separation

Long-delay multipath pushes underwater acoustic mixing beyond the instantaneous model assumed by many classical algorithms; spectral overlap among mechanically and biologically generated sources compounds the difficulty, and low signal-to-noise ratios erode the higher-order statistical cues used by methods such as FastICA and JADE. This work adapts frequency-domain second-order decorrelation (FSD) to convolutive underwater mixtures by using multi-block joint diagonalization of cross-power spectral density matrices in the short-time Fourier transform domain together with compact time-domain regularization of the demixing filters. To provide a controlled and traceable evaluation, we introduce ShipsEarBSS, a simulated benchmark that combines single-source ShipsEar recordings with deep-water BELLHOP arrival responses to form virtual multichannel mixtures with known reference sources. Under a five-trial, eight-SNR protocol spanning −5 to 30 dB, an optimized compact FSD configuration is evaluated against the frozen reference FSD, PCA-SVD, and AuxIVA, and its main design choices are further examined through filter-length, multi-block CPSD, and output-ordering ablations. The results support a cautious conclusion: under the tested ShipsEarBSS protocol, compact time-domain regularization improves the FSD operating point, while the choices of filter support, CPSD block count, and output ordering remain empirical configuration decisions rather than universal optima.