This work develops a fully automatic photogrammetric approach for measuring soil surface roughness from pictures taken in the field with a simple digital camera, without geometric constraints. On each site, 13 overlapping photographs of the soil surface were taken from different angles, under the shade of an umbrella. Millimeter accuracy 3D soil models were calculated from these pictures and were used to derive 11 roughness indexes. The whole procedure was implemented in a fully automatic Python program. The system accuracy was determined on artificial models built with polystyrene, the positional and elevation accuracies of which were about 1.5 mm, while the error on the surface area estimation was less than 0.76% of the site surface area. This approach was successfully applied to an agricultural field experiment in which four soil tillage levels have been generated. These levels were correctly identified using two indices for 96% of the 32 measurement sites. These results show that two roughness indices, the surface tortuosity index and the mean value of height, are most efficient to discriminate agricultural soil tillage levels.