TSNet with sigma-adaptive spectral sampling for tea moisture monitoring
摘要
Accurate, non-destructive estimation of leaf moisture during tea withering is foundational for closed-loop control and quality assurance in industrial tea processing. Although near-infrared spectroscopy is widely adopted for this purpose, field deployment remains fragile: leaf folding/occlusion and pose drift introduce spectral instability and necessitate frequent recalibration, undermining robustness and scalability. To overcome these limitations, we develop SpecTea, an end-to-end system that integrates robust spectral acquisition, sequence modeling, and real-time inference. The method exploits the dynamics of withering to enhance spectral separability and reduce information entropy, and introduces a SASS strategy that combines spatially adaptive five-point segmentation with dynamic Poisson-disk sampling to select informative regions. A custom spectroscopic apparatus secures representative measurements, which are modeled by TSNet, a sequence architecture that fuses self-attention with GRUs for moisture prediction. On held-out tests, TSNet achieves a correlation coefficient of 0.9917, an RMSE of 2.7612, and an RPD of 2.7552, surpassing traditional machine-learning baselines and recent deep models. The SASS-enabled acquisition consistently reduces inter-spectral variance, improving generalization across batches and processing conditions. When deployed on the withering line, the system attains an average error of 3.89%, comfortably meeting the