Patch size

The size of the overlapping patches is a function of the non-stationarity of the data. Smaller patches represent rapidly changing data better but are more expensive and likely to match small patterns of correlated noise. For the examples on multiples we used patches that were just a few samples long (less than 10 in all axes). For the ground-roll example, we used patches 200 samples long in the time axis and 80 samples long in the offset axis.