An iterative, physically model based inversion algorithm has been used to estimate root mean square (r.m.s.) surface roughness height from radar data. collected over vegetated areas. The model is based on a discrete scatterer random media technique, and employs the distorted Born approximation to model the backscatter coefficients for a given scene. In the model, the Fresnel reflectivity (a measure of soil moisture) and surface roughness appear together in the vegetation-ground interaction term. An approach is followed that utilizes differences in their frequency response to separate the two. Sensitivity analysis shows that the change in surface reflectivity owing to the change in frequency from the L- to C-band is dominated by surface r.rn.s. height. The Fresnel reflectivity stays almost constant over this frequency interval. The inversion algorithm based on these sensitivity differences is applied to the backscatter model data from a plant canopy of soybean. Calculations show that the technique gives accurate results from a model backscatter data set that is corrupted with random noise. The inversion algorithm is also applied to Synthetic Aperture Radar (SAR) data collected over corn fields during the MACHYDRO'90 experiment in Pennsylvania, USA. There is an excellent agreement between the measured and the estimated r.m.s, surface height.