STAMPformer: spectral-prior-guided multi-view attention learning for hyperspectral prediction of continuous frost damage severity in field-grown tea plants

Frost damage severely constrains the growth, development and quality formation of tea plants, underscoring the need to develop accurate field-scale methods for the assessment of frost damage severity. In this study, hyperspectral data and eight key physicochemical indicators were collected from tea plants across natural field frost damage gradients. To characterize frost damage severity on a continuous scale, a comprehensive tea cold stress response index, termed TCSRI, was developed by integrating multidimensional physiological responses. To enable the accurate and non-destructive prediction of physicochemical indicators and TCSRI under high-dimensional and small-sample conditions, we propose STAMPformer, a spectral-prior-guided multi-view attention pooling model. It incorporates a parameter-free attention convolution extraction module to jointly model raw spectra and first-order derivative spectra, thereby enhancing local spectral features and suppressing noise. A multi-view attention pooling module further adaptively aggregates sequence features from different feature subspaces to strengthen key spectral responses. The experimental results showed that STAMPformer achieves the best performance across all nine prediction tasks, with R 2 values above 0.87 and RPD values above 2.80. Ablation experiments, paired tests with multiple random seeds, and SHAP analysis confirmed the contributions of the proposed modules and reveal physiologically meaningful key wavelengths associated with pigment degradation, photosynthetic impairment, and cellular structural disruption. Independent external validation further demonstrated the model’s predictive ability on unseen samples. Overall, the proposed TCSRI and STAMPformer provide an accurate, interpretable and non-destructive framework for assessing tea plant frost damage severity, supporting frost monitoring, cold-tolerant germplasm evaluation, and smart tea garden management.