Detection of Moisture Content of Tobacco Materials Based on Hyperspectral Imaging and an Improved CNN-LSTM Model

Highlights An improved CNN-LSTM model is proposed, integrating a multiscale CNN with a bidirectional LSTM for hyperspectral moisture prediction in tobacco. The model achieves superior accuracy (R2p = 0.9931, RMSEP = 0.0037), outperforming conventional machine learning and deep learning methods. SHAP analysis identifies key moisture-sensitive bands (1180 to 1230 nm and 1350 to 1450 nm), providing physicochemical interpretation for feature selection. Pixel-wise moisture visualization enables spatial distribution mapping, supporting online detection and quality control in tobacco processing. Abstract. Accurate determination of moisture content is essential in tobacco processing, as it directly influences product quality and process stability. Hyperspectral imaging provides a rapid and nondestructive solution for moisture detection; however, the high dimensionality, strong inter-band correlations, and limited interpretability of existing models remain major challenges. To address these issues, this study proposes an improved CNN–LSTM model for predicting moisture content in tobacco materials using hyperspectral data. The proposed approach integrates multiscale CNN networks for local spectral feature extraction with an LSTM network to capture sequential inter-band dependencies. Prior to modeling, spectral data are preprocessed using Savitzky-Golay smoothing and 3σ-based outlier filtering to reduce noise and improve data quality. The performance of the proposed model is systematically compared with classical machine learning methods (MLR, PLSR, SVR, and DTR) and conventional deep learning models (RNN and CNN). The results demonstrate that the CNN–LSTM model achieves superior predictive performance, with a test set R²p of 0.9931 and an RMSEP of 0.0037, outperforming all comparative models and existing approaches reported in the literature. To enhance model interpretability, SHAP analysis is employed to quantify wavelength-level contributions, revealing two dominant spectral regions (1180 to 1230 nm and 1350 to 1450 nm), with the latter showing a stronger influence on moisture prediction. This work provides a reliable, interpretable, and visually supported framework for online moisture monitoring in tobacco processing.