A CNN model for predicting soil properties using VIS-NIR spectral data

Abstract

This research aims to develop a novel deep learning-based model for predicting soil properties based on visible and near-infrared (VIS-NIR) spectroscopy data. Soil samples were collected from the European topsoil dataset provided by the LUCAS project provides various soil physicochemical properties analyzed within 28 EU countries (including sand, silt, clay, pH, organic carbon, calcium carbonates (CaCO3), and N). In this study, one-dimensional (1D) convolutional neural network (CNN) models were developed using absorbance spectral data. The performance of feature learning from discrete wavelet transform as a powerful preprocessing method was tested. Moreover, the results of the proposed CNN model were compared with partial least squares regression (PLSR) with raw absorbance and optimum classical preprocessing (Savitzky-Golay smoothing with first-order derivative). The ratio of percent deviation (RPD) of CNN with absorbance data for prediction of OC, CaCO3, pH, N, sand, silt, and clay content were 4.02, 3.89, 2.82, 3.02, 1.63, 1.43, and 2.16, respectively. While the RPD of PLSR with optimal preprocessing of absorbance data for predicting the mentioned parameters were 2.89, 3.00, 2.79, 2.50, 1.37, 1.27, and 1.84, respectively. The study demonstrated the feasibility of using deep learning-based models and VIS-NIR spectral data as a rapid nondestructive tool for the assessment of important soil properties.