Semi-automatic monitoring of soil cover after prescribed burning by image classification using convolutional neural networks

Abstract

We developed a monitoring procedure of the effects of prescribed fire on the soil surface vulnerable to erosion processes based on digital image analysis. We took vertical pictures of the soil surface from a height of 138 cm in 330 sampling points distributed in the diagonals of fifteen 200 m²- plots, including 3 control plots, within 10 ha of a prescribed burnt Pinus pinaster Ait plantation. During the analysis of the images we manually discriminated the surface cover of needles, cones, wood, moss, burnt material, dry and green biomass, rock fragments, and mineral and organic soil. The results showed that the automatic classification of images through convolutional neural networks provides a precise classification of the soil surface and reduces the time required for field data collection and for the identification of the different soil covers, particularly the discrimination of the exposed mineral soil susceptible to erosion. We therefore suggest that broadening and standardizing this methodology may provide significant benefits for monitoring soil surface conditions and erodibility.