Extraction of soil clod properties from high-resolution DEMs using instance segmentation

Soil surface roughness (SSR), which refers to surface irregularities, is a key parameter for determining soil condition and tillage outcomes. Consequently, automatic characterization of soil clod size from images has received considerable attention.

In this study, a dataset of 2.5D digital elevation models (DEMs) acquired in the field at the metre scale was analysed. Clods with diameters as small as 1 cm were segmented using the wavelet-based instance segmentation algorithm proposed by Vannier et al. (2026). The segmented clods were then fitted with half-ellipsoids, and the probability density functions of the resulting shape parameters were estimated.

The three dimensions of each clod were characterized by the lengths of the corresponding semi-axes. In addition, the cloddiness of the entire soil surface was quantified using the two parameters of a cumulative distribution function (CDF) fitted to a semi-axis. The moderate to strong correlations among the three semi-axes suggest that clod size can be adequately described using a single horizontal semi-axis and optionally complemented by the vertical semi-axis. Furthermore, the orientation angles and centroids of the ellipses fitted in the horizontal plane may provide additional insight into the spatial organization of the clods.

Instance segmentation of 2.5D DEMs enables a rich and accurate characterization of soil clod size and organization. Although clod-size classes can be derived if required, the fitted CDF provides a more informative description of the soil surface. This approach may therefore be useful for assessing soil tilth in precision agriculture.