Pros

• The predictive signal/noise separation method operates successfully, even when the noise is severely spatially aliased. The data need not be sampled finely in space to remove ground roll, a fact which is profound, considering the high cost of land-based seismic acquisition. Crawley 1998 shows another example of the feasibility of using PEF's to estimate aliased, coherent seismic signals.
• Ground roll comes in many different flavors. A method which exploits the particular moveout patterns of one observed case of ground roll is sure to fail on others. The predictive method is blind to this; assuming a first order separation gives a viable noise model, the specific moveout patterns of signal and noise are irrelevant.
• Ground roll is almost always dispersive. The f-x-based algorithm of Spitz 1999 is computationally more efficient than our t-x domain algorithm, but it requires a time-invariant seismic wavelet. Ground roll is usually a potpourri of different wave trains, all with different dispersion rates, strongly violating this temporal stationarity assumption. The nonstationary t-x domain technique has no such limitation. Nonstationary filtering is possible in the frequency domain, but the distinct computational advantage over time domain methods is lost.
• An effective, general method for ground removal in 3-D may have profound effects on the future of seismic acquisition. If severe ground roll necessitates the recording of extremely long offsets or the use of large receiver arrays, the ability to remove the ground roll robustly from single-sensor data will cut acquisition costs considerably. Such a ground roll removal technique could greatly assist smaller-scale survey efforts (environmental or university research) which do not boast the same resources as a large, multinational oil company.
• Parenthetically, it should be emphasized that we have presented a method to do signal/noise separation, not simply noise removal. The output estimated noise may contain useful information, such as shear wave velocities. Multicomponent seismology is a promising specialty which would certainly benefit from any additional constraints provided by a robust ``noise'' extraction.
• Considerable effort is currently expended in pursuit of the perfect noise model, particularly in multiple suppression. However the results presented here show that predictive signal/noise separation produces good results with an imperfect noise model. An open question remains: might more primitive methods of obtaining a multiple model suffice if passed to a predictive signal/noise separation algorithm?

Cons:

• As discussed in Appendix A, we are concerned that the current approach may not fully account for correlation between signal and noise.
• While the nonstationary t-x PEF estimation is more robust than stationary f-x methods in terms of accurately predicting all the coherent events, the cost is considerably higher. As mentioned above, the method must be optimized before it becomes an industry standard.
• Our approach is parameter-intensive. Fortunately, in practical cases, the parameter choices are similar for most data gathers in the same survey.