Fusing band and feature attention in CNN-LSTM: A dual-attention framework for Hyperspectral-based precise prediction of nicotine levels in cured tobacco leaves

Rapid and reliable quantification of key chemical constituents in cured tobacco leaves remains a critical challenge in industrial processing. This study introduced an intelligent prediction framework that combines hyperspectral imaging (HSI) with a deep neural network model enhanced by dual-attention modules. A multi-region spectral acquisition strategy was employed to capture representative reflectance data from different leaf areas. To improve feature learning, a Dual-Attention-enhanced CNN-LSTM (DACL) model was developed by incorporating a Band Attention (BA) module for adaptive spectral band selection and a Self-Attention (SA) module for deep feature refinement. The DACL model was trained on multi-region averaged spectra and systematically evaluated against classical machine learning models (PLSR, SVR) and deep learning baselines (CNN, LSTM, CNN-LSTM). The DACL approach achieved superior performance, with an R² of 0.857, an RMSE of 0.443, and an RPD of 2.642 on the independent test set. Visualization of the attention modules confirmed that the model consistently prioritized wavelengths corresponding to nicotine’s characteristic absorption bands, thereby enhancing both predictive accuracy and interpretability. The results demonstrate that the fusion of spectral and deep feature attention within a CNN-LSTM architecture, together with multi-ROI spectral acquisition, provides a robust and interpretable solution for real-time nicotine assessment in cured tobacco. The proposed method offers new insights for quality control in tobacco processing and highlights the broader potential of AI-driven hyperspectral analysis in agricultural product evaluation.