Contemporary data analytics for soil spectroscopy

No soil, no life. Know soil, know life. Soil provides the basis for life. To promote soil security, soil monitoring is essential. However, conventional methods of soil analysis are costly and time-consuming. This thesis explores contemporary data analytics for analyzing soil infrared spectroscopy data. New data analytics take soil infrared spectral data and convert them to soil properties that are useful for assessing its conditions. This thesis deals with issues of sampling, spectral reduction techniques, deep learning models, and application in soil contamination assessment. Soil spectral data has to be trained using machine learning models to provide predictions for soil properties. The effect of sampling size and designs on the model performance were evaluated. Various ways of spectra data dimension reductions were explored using variable selection techniques to prevent model overfitting when a limited number of samples was available. To deal with large data collected from regional and national soil spectral libraries, deep learning techniques were explored. The convolutional neural network (CNN) was demonstrated as a highly accurate method for predicting soil properties on a large database. A method was derived to enable the interpretability of the CNN model. The application of infrared spectroscopy in screening soil contaminants (microplastics and petroleum hydrocarbons) were illustrated. Collectively, this thesis provides novel data analytics that enabled enhanced applications of infrared spectroscopy in soil science.