Prediction of moisture and protein in corn kernels from multiple origins based on NIR-PLSR with gradient boosting machines for feature selection

Zheng, Runyu

Permalink

https://hdl.handle.net/2142/124601 Copy

Description

Title

Prediction of moisture and protein in corn kernels from multiple origins based on NIR-PLSR with gradient boosting machines for feature selection

Author(s)

Zheng, Runyu

Issue Date

2024-05-02

Director of Research (if dissertation) or Advisor (if thesis)

Kamruzzaman, Mohammed

Committee Member(s)

• Allen, Cody M.
• Rausch, Kent D.
• Singh, Vijay

Department of Study

Engineering Administration

Discipline

Agricultural & Biological Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Gradient Boosting Machine (gbm)
• Feature Selection
• Shapley Additive Explanations (shap)
• Partial Least Squares Regression (plsr)
• Corn Kernels
• Near-infrared (nir) Spectroscopy
• Component Prediction.

Language

eng

Abstract

Differences in moisture levels and protein content impact both nutritional value and processing efficiency of corn kernels. Near-infrared (NIR) spectroscopy can be used to estimate kernel composition, but models to do so are typically trained on samples collected from only a few environments which can lead to underestimation of both the error rates and bias of models. In this study, corn samples grown across an internationally diverse set of environments were assembled. NIR spectroscopy with chemometrics and partial least squares regression (PLSR) was used to determine moisture and protein of this international panel of corn grain samples. The potential of five feature selection methods to improve prediction accuracy by extracting sensitive wavelengths for moisture and protein in corn kernels was assessed. SHapley Additive exPlanations (SHAP) values were used to measure the impact of each feature/wavelength on the model prediction. Gradient boosting machines (GBMs), specifically CatBoost and LightGBM, were effective in selecting crucial wavelengths for moisture (1409, 1900, 1908, 1932, 1953, and 2174 nm) and protein (887, 1212, 1705, 1891, 2097, and 2456 nm), producing PLSR models with coefficients of determination of validation (R2V) of 0.97 and 0.82, root mean square errors of validation (RMSEV) of 0.45% and 0.51%, and ratios of performance to deviation of validation (RPDV) of 6.20 and 2.41, for kernel protein and kernel moisture content, respectively. SHAP plots revealed the significant contribution of 2174 nm to moisture prediction and 1891 nm to protein prediction as well as their respective influence tendencies. These results illustrate the effectiveness of GBMs in NIR spectroscopy in feature engineering for predicting chemical components in the agriculture and food sectors, including developing a multi-country global calibration model for moisture and protein in corn kernels.

Graduation Semester

2024-05

Type of Resource

Text

Handle URL

https://hdl.handle.net/2142/124601

Copyright and License Information

Copyright 2024 Runyu Zheng

Owning Collections