Towards a dynamic soil survey: Identifying and delineating soil horizons in-situ using deep learning

Abstract Soil horizons are the manifestation of pedogenesis and contain the basic morphologic indicators of soil formation. Accurate and quantitative tools for delineating soil horizons in-situ can assist soil scientists towards rapid and dynamic soil survey information. The objective of this study was to develop a deep-learning-based, soil-profile-imaging method for identifying and delineating soil master horizons to assist in digital soil descriptions. A total of 160 soil profile images from four soil orders (Alfisols, Entisols, Inceptisols, and Mollisols) were collected from north China in the Inner Mongolia and Liaoning regions. The 160 profile images were amplified to 2400 individual images for model building using data-augmentation procedures. The augmented profile image dataset was divided into training (70%), validation (15%), and test (15%) datasets. The training and validation datasets were imported into a nested, U-net network for model building. The proposed deep-learning (DL) model classified soil profiles into A, B, and C horizons. The mean pixel accuracy of the DL model was 0.86 with the training dataset, 0.82 with the validation dataset, and 0.83 for the test dataset. Results showed that the DL model was judged to be accurate enough for identifying and delineating soil master horizons from profile images in practice. Based on the proposed DL model, a smartphone application was subsequently developed to digitalize the soil profile and assist field evaluation of soil profile horizonation. The smartphone application, for the Android operating system (version 5.0 and greater), featured a response time of