The structure of the soil surface is a key parameter to understanding soil properties and physical processes related to water infiltration or runoff and soil erosion. A seedbed surface has a cloddy structure, which is responsible for its random roughness. The present study focuses on seedbed cloddiness characterization. It relies on a new gradient-based method, introduced recently in order to identify automatically the individual clods or large aggregates on a three-dimensional digital elevation model (DEM) of the soil surface. We analyze the sensitivity of clod identification and characterization to the gradient estimation method. For this purpose, the partial derivatives are evaluated by four different methods, and a ground truth of clod boundaries has been made by a soil scientist on two seedbed surfaces (an artificial surface and an in-field one). The gradient method involving an error in the fourth power of the spacing and the greatest number of neighboring pixels in the estimation enables an improvement in the sensitivity of clod identification by up to 10 percent, achieving detection performance greater than 73 percent. Comparing segmented contours with ground truth delineation, good agreement is found for half of the clods regardless of the numerical gradient computation method. For the other half, the area is underestimated because it is not possible to extend the contours. The mean variation of shape parameters is slightly sensitive to the gradient computation, with differences of around 5 percent. When decreasing the DEM resolution from 1 mm to 2 mm, the difference between the four methods of gradient computation is strengthened.