Deep learning method of visible and near-infrared spectroscopy considering small sample learning and prediction uncertainty assessment to detect soluble solids content of mandarin

The objective of this study is to propose a deep learning (DL) model of Vis-NIR spectroscopy considering the small sample learning (SSL) and prediction uncertainty assessment to detect the SSC of mandarin, called SSL-UncertNet model. Discrete wavelet transform (DWT) module for spectral feature enhancement and multi-head uncertainty quantification (MHUQ) module for spectral prediction uncertainty assessment were combined with hierarchical adaptive residual network to construct the SSL-UncertNet model. Ablation experiments based on cross-validation performance demonstrated the effectiveness of DWT and MHUQ modules for SSC prediction. Additionally, a model fine-tuning strategy with prediction uncertainty significantly enhanced the model generalization performance for SSC prediction on test set, exhibiting an R 2 of 0.961, an RMSE of 0.264, and a MAE of 0.196. The SSC prediction performance of proposed models also outperformed traditional machine learning models and existing related researches. Overall, the SSL-UncertNet model of Vis-NIR spectroscopy for detecting mandarin SSC provides an effective solution for the generalization limitations of the SSL and spectral prediction uncertainty. • Considering sample sample learning and prediction uncertainty assessment. • Developing the SSL-UncertNet of Vis-NIR spectroscopy for SSC prediction. • DWT and MHUQ exhibited the effective effects on for SSC prediction. • Model fine-tuning strategy improved the generalization ability.