In marine simultaneous source acquisition, although acquisition efficiency has significantly improved, the nearly concurrent firing of multiple sources introduces strong interference among source wavefields in the recorded data, which must be deblended before further processing. In this study, we adopt an iterative shrinkage thresholding deblending framework in a sparse transform domain. For such methods, designing an appropriate thresholding strategy to handle transform domain coefficients effectively is critical. To this end, the nonsubsampled contourlet transform (NSCT), which provides a multiscale, multidirectional, and shift-invariant representation, is employed. A refined shrinkage thresholding strategy tailored to the NSCT domain coefficients of blended seismic data is proposed. First, the blended data are decomposed into subbands at different scales and directions using the NSCT. Subsequently, an adaptive threshold estimation scheme is developed to accommodate the different distributions of useful signal and blending noise across scales and directions. However, coefficient heterogeneity exists not only across NSCT scales and directions but also within each directional subband. Therefore, after constructing scale- and direction-adaptive baseline thresholds, a twin support vector machine classifier is introduced to distinguish useful-signal-dominant coefficients from blending-noise-dominant coefficients, enabling class-dependent threshold refinement. Finally, shrinkage is applied to the NSCT coefficients using the refined coefficient-wise thresholds. Tests on synthetic and field datasets verify the effectiveness of the proposed method in separation.