On-site identification of black soil thickness based on drill-core imaging and deep learning

Abstract

Accurate identification of the black soil thickness from soil profiling is usually time-consuming and labor-intensive, while the on-site identification of black soil thickness by experts is challenging due to the notable transition zone in the thick black soil horizon. This study proposes a framework for efficient identification of black soil thickness from drill core imaging using smartphone and deep learning. Without excavating a soil profile, drill core images from a carry-on soil sampler can be used to identify the black soil horizon using a trained deep learning model of the VGG-16 backbone U-net algorithm. The approach was tested with a limited dataset obtained from field sites in the black soils of northeast China and the results show that it can efficiently identify the black soil horizon on site. A good accuracy was obtained, with R² = 0.95 and RMSE = 0.07 m for the estimates of black soil thickness. Overall, the proposed methodology offers the possibility of efficiently identifying black soil thickness on a large scale, thus accurately quantifying regional black soil degradation.