Spectral-Based Machine Learning Enables Rapid and Large-Scale Adsorption Capacity Prediction of Heavy Metals in Soil

Abstract

Accurate and large-scale estimation of the soil adsorption capacity of heavy metals (HMs) is vital to tackle soil HM contamination. Here, a novel framework has been developed to evaluate the adsorption capacity of HMs in soil using visible and near-infrared spectroscopy. Soil attributes were accurately estimated without any spectral preprocessing using a combined autoencoder (AE) and deep neural network (DNN) approach. Soil HM adsorption capability was then evaluated based on spectral-derived soil attributes, using 2,416 data points on Cd(II), Pb(II), and Cr(VI). The proposed AE-DNN models offer accurate estimations of soil attributes with an average R² of 0.811 on the independent testing sets. The trained AE-DNN models can reveal patterns typically used by experts to identify bond assignments and promote data-driven knowledge discovery. By comparison with adsorption capacity maps based on actual and estimated soil attributes, we show that the spectral-based soil adsorption capacity evaluation is statistically reliable. Our adsorption capacity maps for the EU and USA identify known soil contamination sites and undocumented areas of high contamination risk. Our framework enables rapid and large-scale prediction of the adsorption capacity of HMs in soil and showcases important guidance for further soil contamination testing, soil management, and industrial planning.