Rapid and non-destructive detection of hard to cook chickpeas using NIR hyperspectral imaging and machine learning

Abstract

Identification of Hard to Cook (HTC) chickpeas in a rapid, non-destructive manner is crucial for the pulse processing industry. This study investigated the potential of near infrared (NIR) hyperspectral imaging (HSI) system to classify chickpeas into HTC and Easy to Cook (ETC) (control) categories. Two types of HTC chickpeas were created using eight different varieties of chickpeas: the first type was created by storing under suboptimal conditions, while the second type was created with chemical treatment. A total of eight hundred sixty-four chickpea seeds ({control- 36; physically hardened-36 seeds; chemically hardened-36 seeds} × 8 varieties) were used in this study. The chickpeas were imaged using a NIR-HSI system in the spectral range of 900–2500 nm. The cooking time of individual chickpea seed was measured using an automated Mattson cooker and the spectral data was correlated with the measured reference cooking time of chickpeas to develop the calibration model. Partial Least Square Discriminant Analysis (PLSDA), Support Vector Classifier (SVC) and Convolutional Neural Network-Attention (CNN-ATT) models was used for model development based on full spectrum and significant wavelengths. The optimal models were obtained using the SVC and CNN-ATT which demonstrated 100% accuracy in classifying the chickpeas into HTC and ETC. Besides, the cooking time of control (ETC) and HTC chickpeas were predicted using One Dimensional Convolutional Neural Network (1D-CNN) with Correlation Coefficient of Prediction (R2p) and Root Mean Square Error of Prediction (RMSEP) values of 0.880 and 0.662 respectively indicating the potential of this approach in developing robust model for cooking time prediction in other pulses.