Attention-infused 1D-CNN driven by near-infrared spectroscopy for robust moisture prediction across the full chain of morphologically heterogeneous green broken-tea processing

In the rolling and four - stage drying processes of green broken tea, pronounced morphological heterogeneity and process coupling lead to substantial scattering effects, which complicate the real - time monitoring of moisture content throughout the entire production chain. This research presents an attention - 1D - CNN model that integrates dual - path channel - spatial attention mechanisms. Leveraging near - infrared spectroscopy, a continuous spectral - moisture dataset consisting of 720 samples, spanning from fresh leaves to the final drying stage, was established. By introducing a learnable dynamic attention mechanism, the model effectively suppresses scattering noise and adaptively enhances the features of moisture - sensitive spectral bands. Ablation experiments and independent validation indicate that the model attains an R² value of 0.9323, a root - mean - square error (RMSE) of 0.0462, and a mean absolute error (MAE) of 0.031 on the test set, outperforming partial least squares (PLS), support vector regression (SVR), and conventional 1D - CNN by 8–11%. In the rolling and four - stage drying phases, where morphological heterogeneity is most prominent, the prediction error is reduced by over 24%, demonstrating the model's robustness against the coupled effects of morphology and process. The visualization of attention weights further validates its adaptive focus on moisture - sensitive spectral regions, surmounting the limitations of traditional global modeling and manual feature selection. This study not only offers a practical deep - learning approach for online moisture detection in multi - stage tea processing but also provides a novel methodological reference for the spectral analysis of highly heterogeneous agricultural and forestry products.