Machine Learning Analysis of Raman Spectra To Quantify the Organic Constituents in Complex Organic-Mineral Mixtures

Area of Science:

• Agricultural Science
• Analytical Chemistry
• Spectroscopy

Background:

• Soil chemical composition is crucial for agricultural policy and practice.
• Raman spectroscopy offers rapid, non-invasive analysis of organic matter but faces challenges in soils due to complex composition and fluorescence.
• Overcoming spectral interference is key for accurate soil analysis.

Purpose of the Study:

• To compare the performance of convolutional neural network (CNN) and partial least-squares regression (PLSR) models for analyzing amino acid composition in soils.
• To evaluate the impact of spectral sampling and data preprocessing on model accuracy.
• To identify optimal modeling strategies for Raman spectroscopy of soils with mineral and fluorescence interference.

Main Methods:

• Created soil-mimicking mixtures of amino acids, fluorescent bentonite clay, and mineral components.
• Applied and compared CNN and PLSR multivariate models to Raman spectral data.
• Utilized volume-averaged spectral sampling and data preprocessing techniques.

Main Results:

• Both PLSR and CNN models demonstrated high predictive accuracy (R-squared values up to 0.98 for PLSR and 0.97 for CNN).
• Performance was influenced by matrix complexity and signal-to-noise ratio.
• CNN models outperformed PLSR in scenarios with very low organic signal-to-noise ratios.

Conclusions:

• Volume-averaged sampling and data preprocessing significantly enhance Raman analysis of soils.
• PLSR and CNN models offer comparable, robust performance for soil organic matter quantification.
• CNN models provide an advantage for analyzing complex soil samples with low organic signal-to-noise, crucial for accurate agricultural applications.