Hyperspectral imaging coupled with transformer enhanced convolutional autoencoder architecture towards real-time multi-target classification of damaged soybeans

Developing real-time multi-target detection method of soybean quality is a challenging work. This study proposed an innovation strategy that integrating hyperspectral imaging (HSI) and transformer-convolutional autoencoder (T-CAE) with deep learning models to simultaneously identify different damaged soybeans, including broken, infested, diseased, and moldy types. The 1D convolutional neural network (1D-CNN) models based on the T-CAE augmented spectrum can discriminate these different damaged soybeans, with the classification accuracy (Acc) over 96.00 %. Introducing the self-attention mechanism of T-CAE architecture can be a considerable spectral data augmentation strategy to overcome the sample limitation for classification model development. Notably, T-CAE-generated spectra exhibited strong capabilities in capturing global spectral features while preserving local details via convolutional encoding. When used to train 1D-CNN models, they improved Acc by 9.71 % compared to models trained on raw spectrum. Consequently, HSI technique coupled with T-CAE can be a rapid, real-time and multi-target classification strategy for damaged soybeans. • A novel T-CAE strategy was proposed for spectral data augmentation under limited samples. • The strategy integrates hyperspectral imaging with transformer-enhanced convolutional autoencoder. • T-CAE-augmented spectra significantly improved 1D-CNN classification accuracy (>96 %). • The approach enables real-time, multi-target detection of broken, infested, diseased, and moldy soybeans.