Soil surface roughness (SSR), referring to surface irregularities, is a key parameter for assessing soil condition and tillage outcomes. Characterizing roughness at fine scales—including clods and depressions—remains challenging for 2.5D digital elevation models (DEMs) collected at the meter scale in the field. This study presents two segmentation methods for high-resolution DEMs from an agricultural site. For clod segmentation, a wavelet-based approach from the literature was used, while a novel histogram-based method was introduced for depressions. Both methods were evaluated on natural soil surfaces with varying roughness levels and a simulated surface, with and without noise, using standard metrics (recall, precision, F1-score, IoU). The best clod segmentation results were achieved on fine seedbeds (95.2% recall, 97.3% precision, 96.2% F1-score), with slightly lower but strong performance on plowed surfaces (84.2% recall, 96.9% precision, 90.1% F1-score). Due to their lower frequency, depressions were primarily assessed visually under field conditions. For the simulated surface (with ground truth), IoU values ranged from 84.2% to 87.9% for clods and around 92% for depressions, demonstrating competitive performance. Additionally, the volume of roughness features was computed and visualized using cumulative distribution functions. These segmentation methods enable monitoring of soil surface conditions, with applications in precision agriculture, surface-water interactions, and meter-scale microwave remote sensing.
Vannier et al. (Thu,) studied this question.