Hyperspectral imaging combined with machine and deep learning for non-destructive detection of defective red pepper ( capsicum annuum L.) powder adulteration

This study presents a non-destructive approach for detecting adulterated defective red pepper powder (DRPP) using short-wave infrared (SWIR, 900–1700 nm) hyperspectral imaging (HSI) with advanced preprocessing, dimensionality reduction, and both machine learning and deep learning techniques. Two-dimensional correlation spectroscopy reduced the spectral variables from 112 to 15 informative wavelengths. Among the machine learning models, support vector regression with standard normal variate and Savitzky – Golay first derivative (SG-1) preprocessing achieved the best performance, with a residual predictive deviation (RPD > 7.9; values above 3 indicate excellent predictive ability). Recurrent neural networks with normalization and SG-1 preprocessing also achieved high accuracy (Rc2 and Rp2 > 0.98, where values close to 1 reflect a strong correlation between predicted and actual values). Despite an 86% reduction in spectral data, models based on selected wavelengths maintained or even improved the accuracy of full-wavelength models, demonstrating the effectiveness of the wavelength selection strategy. Additionally, HSI enabled the visualization of the distribution patterns of DRPP samples with diverse degrees of adulteration. Overall, this study confirms the feasibility of SWIR HSI integrated with optimized data processing for real-time, accurate detection of DRPP adulteration, offering a scalable and practical solution for food quality monitoring and adulteration assessment in agricultural products.