深度学习模型在可见-近红外光谱分析中的应用进展OA

Visible-near-infrared(Vis-NIR)spectroscopy is characterized by high dimensionality,nonlinearity,and background interference,yet its analytical capabilities have been significantly enhanced by the introduction of deep learning(DL)models.This paper systematically reviews the advances in the application of DL from 2018 to 2025,outlining the fundamental mechanisms of models such as autoencoders(AE),long short-term memory networks(LSTM),convolutional neural networks(CNN),deep belief networks(DBN),and Transformers,as well as their use in Vis-NIR spectral analysis.In the field of food inspection,Vis-NIR spectroscopy combined with CNN and its variants enables nondestructive quality testing and online sorting,while attention mechanisms and Transformers improve model interpretability to address the ″black box″ problem.For environmental monitoring,where Vis-NIR spectral features are weak and heavily interfered,1D-CNN has successfully extracted spectral characteristics for quantitative and qualitative analysis.Through multi-source data fusion,data augmentation,and hybrid transformer architectures,model generalization and the inversion accuracy of soil and water parameters have been enhanced.In agricultural and forestry monitoring,faced with noisy spectra and scarce samples,integrated approaches using data augmentation,feature selection algorithms,and 1D-CNN have achieved species classification,environmental stress monitoring,and disease screening.In mineral product analysis,to overcome the challenges such as spectral similarity among different materials(same spectrum for different objects)and nonlinearity,1D-CNN has improved lithological identification and mineral property inversion,with interpretability methods further increasing model transparency.Additionally,Vis-NIR spectroscopy combined with DL has been successfully applied in areas such as dating ancient ceramics,plastic sorting,and vitality detection of poultry eggs.Looking forward,DL models in Vis-NIR spectroscopy will see integrated innovations in model architecture,data augmentation,and interpretability.By promoting standardization and normative application,this technology is expected to develop toward greater reliability and broader adoption.