Degree
Doctor of Philosophy (PhD)
Department
The Gordon A. and Mary Cain Department of Chemical Engineering
Document Type
Dissertation
Abstract
Sugarcane is a critical agricultural commodity, supporting sugar production, biofuel industries, and rural economies worldwide. The global sugar industry is valued at approximately 78 billion dollars annually. Cane payment systems between factories and growers in Louisiana and other major sugarcane-producing regions rely on quality parameters such as pol, Brix, and moisture, traditionally determined through wet chemistry methods. Although effective, these methods are labor-intensive, time-consuming, and vulnerable to variability. Rising levels of extraneous matter (EM), including soil and leaves, have further impacted sugar recovery, driven by climate change, mechanized harvesting, and green cane regulations. Despite its influence, EM content is not routinely measured in factories due to the lack of practical methods. This challenge is particularly critical in Louisiana, where EM levels are among the highest reported globally, potentially affecting calibration model performance.
This dissertation evaluates the application of Near-infrared (NIR) spectroscopy combined with machine learning (ML) models as a rapid and nondestructive alternative for analyzing sugarcane quality parameters and EM content. Calibration models for Brix, pol, moisture, and EM were developed using partial least squares regression (PLSR), functional regression (FR), support vector regression (SVR), and artificial neural networks (ANN). Total leaf content was predicted from shredded cane mixtures with known concentrations of cane, leaves, and soil, while soil content was predicted based on incinerated ash as the reference method.
Machine learning (ML) models achieved strong performance across most parameters, with coefficients of determination (R²) typically exceeding 90 percent and root mean square errors (RMSE) remaining below 10 percent of the measured range. FR consistently provided improved performance compared to PLSR. Improving prediction accuracy is critical because NIR predictions are directly tied to payment calculations, and any errors can affect financial transactions between growers and factories.
This research presents the first comprehensive methodology to quantify EM in sugarcane using NIR spectroscopy, addressing a longstanding gap in quality evaluation. The approach not only advances sugarcane industry practices but also offers potential applications for other crops where EM affects processing efficiency and product value. These findings support the future integration of NIR and ML models into factory operations and payment systems, promoting more accurate and economically sustainable agricultural industries.
Date
6-12-2025
Recommended Citation
Imbachi Ordonez, Stephania, "Rapid Quality Assessment of Sugarcane Using Near-Infrared Spectroscopy and Machine Learning Models" (2025). LSU Doctoral Dissertations. 6817.
https://repository.lsu.edu/gradschool_dissertations/6817
Committee Chair
McPeak, Kevin M.