In terms of simplicity a TIN can provide a very reasonable return.
I see no reason why noise precludes the use of a TIN per se. They will exactly model your data points if you set the parameters up that way or fit more smoothly to the surface. They also have the distinct advantage of being scale and grid orientation independent, unlike any moving window based method.
I would suggest :
- Check WolfOdrade's point first!
- If the bathymetry data is arranged in rows or some other irregular geometry (e.g. from boat soundings) then topogrid these these to a raster surface. If it is points in a regular grid or randomly scattered, make it into a TIN that fits every point.
- Visualise in something simple, like ArcScene.
- Then Post a picture of it somewhere so we can actually see the data, showing your points of concern.
The right smoothing algorithm is very dependent on the type of terrain, Guassian filter might be good for smoother DEM. FFT would be worth the effort for varied terrain and could handle sharper angles, but iteratively weeding the TIN nodes could be a sufficient and simpler option that minimises simulating data.