Deep automatic soil roughness estimation from digital images

Abstract

Soil roughness, defined as the irregularities of the soil surface, yields significant informationabout soil water storage, infiltration and overland flow and, thus, is a key factor in characteriz-ing the quality of the terrain; it is often used as input in many synthetic general agriculturalmodels and in particular in soil moisture estimation models. In this paper, we proposea framework that combines a specific setup for data acquisition with deep convolutionalnetworks for actual estimation. The former relies on projecting a line red laser beam on theanalysed soil surface followed by digital color image acquisition. The later, involves twoconvolutional models that are trained in a supervised manner to predict the soil roughness.The data set was produced in the laboratory both on synthetic and real soil samples. The labelsused in the training process are the soil roughness values measured by using a pinboard. Thedetailed evaluation showed that the error of the automatic precision lies in the range of groundtruth deviation, thus validating the proposed procedure.