Glucosinolate quantification in multi-cultivar broccoli powder using explainable deep learning

Glucosinolates (GSL) are bioactive phytochemicals with discernible health and nutritional value. Rapidly quantifying GSL in broccoli powder is crucial for food processing and crop breeding. This study developed a predictive model for total GSL content in freeze-dried broccoli powder by integrating multispectral imaging technology with chemometric methodologies. Spectral data and image texture features were fused to compare the performance of three deep learning architectures—convolutional neural network (CNN), CNN-bidirectional long short-term memory (CNN-BiLSTM), and CNN-BiLSTM with attention mechanism (CNN-BiLSTM-ATT). Five-fold cross-validation was used to evaluate model robustness. The CNN-BiLSTM-ATT model, coupled with Savitzky-Golay (SG) smoothing and standard normal variate (SNV) preprocessing, achieved optimal performance (R 2 TEST = 0.907 ± 0.018, RMSE TEST = 2.086 ± 0.291 μmol/g). Additionally, SHapley Additive exPlanations analysis elucidated the contribution of fused information to model accuracy. This high-throughput detection tool enables precise GSL profiling in broccoli powder, supporting advanced applications in powder-optimized cultivar screening and functional food manufacturing. • Deep learning enhances multispectral imaging for glucosinolate prediction. • CNN-BiLSTM-ATT model outperforms CNN and CNN-BiLSTM architectures. • Image texture information can complement spectral data for glucosinolate detection. • SHAP analysis provides insight into feature contributions. • Model supports rapid, high-precision glucosinolate detection in food.