Simultaneous quantification of multiple chemical properties of soil solution using smart spectroscopy

Soil environmental monitoring is crucial for crop production, urban planning, and human health. However, the existing monitoring methods are inefficient and costly. In order to improve the efficiency of soil monitoring, a multi-indicator concentration detection method based on a new modified spectrometer technology (MST) was proposed.

MST, a method that combines high-throughput experiments (HTE) and machine learning (ML), was proposed to determine various substances in a complex environment and exhibited features of large measurement throughput and high prediction accuracy. Soil is a classic complex chemical system in nature, so we want to try to apply MST to soil monitoring projects. In this study, about 14,400 holographic scattering spectroscopy (HSS) images were captured using the MST and used to train 3 deep neural networks (ResNet-50, Inception V1, and SqueezeNet V1.1). These models were optimized by adjusting parameters and hyperparameters.

The concentration prediction model based on ResNet-50 has fast convergence speed and a good learning effect. The model can simultaneously detect eight indicators. The best evaluation results achieved coefficient of determination (R²) = 0.996, root mean square error (RMSE) = 0.758, and mean relative error (MRE) < 5% in the test set.

The results show that MST is superior to the previous studies published by EPA, and demonstrate some potential in introducing ML to soil environmental monitoring.