Machine learning of clod evolution under rain for numerical simulation of microtopographic variations by clod layout

Soil surface roughness (SSR) is shaped by tillage operations and evolves with weathering. It is related to geomorphologic processes and to soil fertility. Thus it is an input of various models and the object of many studies. Numerical generation of rough surfaces is an alternative to measurement, which can be cumbersome. This paper proposes a way to numerically generate soil surfaces resembling seedbeds with roughness that have evolved due to rain. As SSR is related to clod-size distribution, the principle is to set modelled clods on a planar surface, as a first approximation of the surface. An experiment was designed to get controlled surface roughness by setting pre-sieved clods on a nearly horizontal surface of loose soil, and then subjecting the surfaces to rainfall events performed by a rainfall simulator. Digital elevation models (DEMs) of each state of the surfaces were recorded by laser scanner to monitor the evolution of clods under rain. Clods were segmented and matched to form a data base of individual clods at each state of the surface. The evolution of clods under rain was modelled by Machine Learning. A set of DEMs was used for learning, the other for test, and several metrics were applied. Obtained results show the robustness of the model. This simple surface representation captures the main properties of the surface. Small scale surface is useful for various applications, such as rough surface scattering, and more generally, modelling where soil surface is an input. • Numerical simulation of a seedbed-like surface evolution under rain is presented. • DEM recording and Machine Learning allowed to learn clod evolution under rain. • Surface representation by clod layout captures the main properties of the surface.