Infrared spectroscopy combined with deep learning to describe the textural properties of cooked rice from raw materials: revealing spectral variations and internal correlations during processing

Abstract

This study investigates the potential to describe the textural properties of cooked rice directly based on their infrared spectroscopy of raw materials, including hardness, adhesiveness, cohesiveness, springiness, gumminess, and chewiness. Near-infrared (NIR) and mid-infrared (MIR) spectroscopy were collected from a large variability of 122 rice varieties in Asian region. The analysis of spectral variance highlighted the thermal processing induced intensive variations of NIR wavelength at 1380 nm and MIR wavenumbers at 890 cm⁻¹. Furthermore, specific spectral regions around 2000 nm and 980 cm⁻¹ showed strong correlations during rice cooking, associated with starch and moisture changes. Convolutional neural networks models based on the NIR and MIR spectrum of cooked rice can satisfactorily predict their textural properties, particularly the hardness with R_v² of 0.92 and 0.95, respectively. Notably, support vector machine models based on the selected MIR and NIR spectral variables of raw materials can directly describe the texture of cooked rice, with the R_v² ≥ 0.90. These results demonstrate that infrared spectroscopy combined deep learning to describe the textural properties of cooked rice from raw materials.