130: Predicting Soybean Quality Based on Hyperspectral Imaging and Machine Learning

Introduction

Predicting soybean quality is critical for ensuring the efficiency of agricultural production and meeting market demands. The standard hyperspectral imaging (HSI) approaches for predicting soybean quality often overlook how physical and environmental factors—such as light penetration depth, soybean surface shape, measurement height, and seed density—can significantly alter spectral signals. The main objective of this study is to explore the depth of light penetration into soybean kernels, as well as to investigate how different factors such as angle, height, and density affect the spectral characteristics, thus evaluating their influence on the prediction models’ performance. This study explores the potential of HSI combined with machine learning to assess soybean quality with enhanced accuracy.

Methods

Hyperspectral imaging data (1000 ~ 2500 nm) of 100 soybeans were collected under controlled experimental conditions. The depth of light penetration was studied by placing soybeans of different thicknesses on paper uniformly coated with quantum dots. Changes in spectral characteristics were observed by varying the angles, density, and height. Preprocessing techniques, including white dark calibration and masking, were applied to the hyperspectral data to enhance its quality. Then, quality parameters like moisture, protein, and ash content were measured, respectively. Subsequently, machine learning algorithms were employed to build predictive models and assess model performance with and without the inclusion of these factors.

Results

The results showed that light can penetrate the seed coat of soybean rather than the whole kernel. Moreover, the surface shape of the soybean had a great influence on the spectral profiles, which were quite different by scanning the curved and flat surfaces of half a soybean. Angles had little effect on predicting soybean quality, however, significant differences were found between the spectra of different heights and densities. Incorporating these factors into the prediction model substantially improves prediction accuracy of moisture, protein, and ash content.

Significance

This study highlights the importance of considering environmental and physical factors when building HSI-based prediction models for soybean quality. By integrating HSI with machine learning while accounting for key influencing factors, this research contributes to advancing precision agriculture, supporting the efficient monitoring and grading of soybean quality.

Authors: Menglin Han, Minwei Xu