Lightweight spatial pyramid pooling convolutional neural network assisted hyperspectral imaging for Hangbaiju origin identification

Abstract

Hangbaiju is a popular food supplement in the form of tea whose quality is influenced by the geographical origin. To control food quality, this work proposed a novel method, hyperspectral spatial pyramid pooling convolutional neural network (HSPPnet) assisted hyperspectral imaging (HSI), to identify the origin of Hangbaiju. HSPPnet utilizes both spatial and spectral information from hyperspectral images effectively and tolerates images of various sizes. This study initially evaluated the performance of four classical classification algorithms (k-NN, RF, XGBoost and PLS-DA) using hyperspectral average spectra, uncovering limitations in distinguishing samples with severe spectral overlap. Additionally, three representative deep learning models (VGG16, ResNet18, and DenseNet121) were used to analyze hyperspectral images compressed into three channels, resulting in overfitting due to the loss of detail information. Finally, HSPPnet and three improved deep learning models (VGG16-25, ResNet18-25, and DenseNet121-25) were used for Hangbaiju origin identification. HSPPnet achieved 100.0% accuracy for both the test set and prediction set, and three improved deep learning models also had been significantly improved. Through comparison, it was found that HSPPnet is an efficient and lightweight model, boasting the benefits of being faster, smaller, and more power-efficient. It was more suitable for deployment on HSI devices to realize large-scale inline detection. The interpretability of HSPPnet was studied to explore the pixel features of Hangbaiju hyperspectral images that were helpful for identification. Additionally, the proposed method exhibited excellent greenness and whiteness properties. The results showed that the proposed method was a non-destructive, efficient, and green method for Hangbaiju origin identification.