It depends on what you mean by 'geometry'. Volume? Mean depth? Surface area? Shape of hypsometric curve?

I assume from the question that your model grid resolution is coarser than the data resolution, in which case the simplest approach would be to form a simple average of observed values within each grid cell.

Averaging has problems if the data distribution is non-uniform. Under such conditions (which you might have from soundings) you have what might be considered to be a classic gridding problem. There is a large literature on such problems, and many possible solutions (Venables & Ripley 1999). If smoothness is desired, you might try using the "Barnes" method, which is used a lot in oceanography (Koch et al. 1983; Barnes 1994).

REFERENCES

The citations given below may be useful, either directly or as entries to the literature.

@book{venables1999mass,

Address = {New York, NY, USA},

Author = {W. N. Venables and B. D. Ripley},

Edition = {Third},

Publisher = {{S}pringer-{V}erlag},

Title = {Modern applied statistics with {S}-plus},

Url = {http://www.stats.ox.ac.uk/pub/MASS4},

Year = {1999}}

@article{koch1983aibo,

Author = {S. E. Koch and M. DesJardins and P. J. Kocin},

Journal = Journal of Climate and Applied Meteorology,

Pages = {1487-1503},

Title = {An interactive {B}arnes objective map analysis scheme for use with satellite and conventional data},

Volume = {22},

Year = {1983}}

@article{barnes1994aobo1,

Author = {S. L. Barnes},

Journal = {Journal of Atmospheric and Oceanic Technology},

Pages = {1433--1447},

Title = {Application of {B}arnes objective analysis scheme. {P}art {I}: effects of undersampling, wave position, and station randomness},

Volume = {11},

Year = {1994}}

I got rid of ripples in my bathymetry with a moving average. Depending on the window size, more or less detail is kept. I used the convolution (http://en.wikipedia.org/wiki/Convolution) function in Python (Numpy module) to do it. You can also do it in Matlab.

My bathymetry was 1D but there are 2D convolution functions in Python and Matlab.