A Quantitative Spectra Analysis Framework Combining Mixup and Band Attention for Predicting Soluble Solid Content of Blueberries

Abstract

Hyperspectral imaging can rapid and non-destructive monitor physical characteristics and intrinsic chemical information of food. In recent years, many studies have applied hyperspectral imaging to evaluate the internal quality of fruits. However, due to the influence of environmental factors, there are abnormal samples in the collected data. Furthermore, the model faces challenges such as limited data availability and insufficient diversity in the dataset. In this study, we collected a total of 1010 hyperspectral images of blueberries and measured their soluble solid content (SSC). To reduce the influence of abnormal samples and increase the diversity of samples, we propose a deep learning framework combining mixup and band attention to predict blueberry SSC. The mixup module performs data augmentation on both spectra and SSC values, enhancing sample diversity and improving the generalization performance of the model. The band attention module captures cross-band information and learns band weights, enabling the model to focus on the bands relevant to SSC. Furthermore, we find that bands with higher weights are consistent with SSC-sensitive bands in existing knowledge, which improves the interpretability of the model.