Integrating NIR Hyperspectral Techniques with Neural Network for Barley Seeds Classification

Nondestructive identification of seed varieties from different origins is crucial in optimizing crop improvement, plant growth, and advancement in breeding methods. Evaluating varietal purity is essential for seed quality. Traditional barley seed classification relies on visual inspection, which is a subjective and error-prone process and is time-consuming. Moreover, chemical testing involves the breaking of seeds which highlights the necessity for exploring non-destructive methods. The integration of Near-infrared (NIR) spectroscopy with neural networks offers a promising alternative. In this study, we have used a hyperspectral imaging dataset consisting of 34 different Indian barley varieties (1008 per variety). Six spectral pre-processing techniques were used to extract and pre-treat the mean reflectance spectrum associated with each seed, namely Savitzky-Golay (SG) smoothing, SG first derivative, SG second derivative, detrending, standard normal variate (SNV), and multiplicative scatter correction (MSC). Further, we have used five different classifiers: K-nearest neighbors (KNN), partial least squares discrimination analysis (PLS-DA), artificial neural networks (ANN), support vector machine techniques (SVM), and convolutional neural networks (CNN). CNN demonstrated superior classification accuracy when implementing it on unprocessed data, whereas the ANN model outperformed it when combined based on the SG2 pre-processing method. Their respective classification accuracy rates were $98.44 \%$ and $98.88 \%$. The current study examined the use of an ANN model coupled with the NIR-HSI technology to accurately, quickly, and nondestructively classify different types of barley seeds.