Research on Detection Method of Moldy Tobacco Leaf Raw Materials Based on Hyperspectral and Deep Learning

In this scholarly pursuit, hyperspectral imaging(HSI) technology was synergistically integrated with advanced deep learning techniques to detect mold contamination in tobacco leaves. The study involved the compilation of a comprehensive dataset of hyperspectral images, which included both mold-infested and healthy tobacco leaves. The spectral data underwent a rigorous preprocessing regime, commencing with Multivariate Scatter Correction (MSC) to mitigate the effects of scattering and followed by Principal Component Analysis (PCA) to isolate the most salient feature bands, culminating in the extraction of representative mean spectral profiles. Subsequently, an array of classification models was meticulously developed, including 1D convolutional neural networks (1DCNN), random forest (RF), and support vector machine (SVM) algorithms. The empirical results conclusively demonstrated the 1DCNN model's superiority in classifying mold-related hyperspectral data, achieving a remarkable training accuracy of 100% and an exemplary testing accuracy of 99.75%. This performance not only surpassed that of traditional machine learning models but also showcased the 1DCNN model's heightened classification accuracy and robust generalization capabilities. These findings herald the potential for deep learning to revolutionize the analysis of complex hyperspectral data, offering innovative solutions to challenges that were previously insurmountable.