Three mechanisms that might produce noise patterns on stacked sections are examined: direct and trapped waves that propagate outward from the seismic source, cable motion caused by the tugging action of the boat and tail buoy, and scattered energy from irregularities in the water bottom and sub-bottom. Different noise patterns can be observed on shot profiles and common-midpoint (CMP) gathers; these patterns can be diagnostic of the dominant mechanism in a set of data. Field data suggest that the dominant noise is from waves scattered within the shallow sub-bottom. This type of noise is enhanced by CMP stacking. Moreover, it can be as strong as surface wave noise on stacked land seismic data as well. Moveout filtering is best for suppressing the noise while preserving signal. Since the scattered noise does not exhibit a linear moveout pattern on CMP-sorted gathers, moveout filtering must be applied either to traces within shot records and common-receiver gathers or to stacked traces. Our data example demonstrates that although it is more costly, moveout filtering of the unstacked data is particularly effective because it conditions the data for the critical data-dependent processing steps of predictive deconvolution and velocity analysis.--Modified journal abstract.