Non-destructive sensing of nitrogen and phosphorus contents in plants via hyperspectral imaging of Hydrangea macrophylla at the seeding stage

Hyperspectral imaging is a powerful, non-destructive tool that has shown promise in estimating plant nutrition but remains underexplored for Hydrangea macrophylla, an important ornamental shrub. In this study, we developed regression and discrimination models for assessing nitrogen (N) and phosphorus (P) statuses in hydrangea at the seeding stage using spectral data combined with machine learning techniques. Spectral reflectance data, captured across various nutritional states following different N and P fertilisation treatments, were paired with laboratory-measured leaf nitrogen content (LNC) and leaf phosphorus content (LPC) data. Spectral reflectance data captured under various N and P treatments were processed with first derivative (FD) and continuous wavelet transform (CWT) to improve the data quality. The results of the correlation analysis revealed that the CWT was more strongly correlated with N (r = –0.90) and P (r = –0.87) than with FD. The machine learning models, which were provided with full-band and wavelet features, showed that the partial least squares regression (PLSR) model, which was integrated with the CWT, could accurately predict LNC (R² = 0.947, RRMSE = 9.6 %) and LPC (R² = 0.827, RRMSE = 10.6 %). Additionally, the PNN model classified nutrient statuses with over 95 % accuracy. Notably, N predictions outperformed those for P, probably because of weaker spectral correlations with P. These findings highlight the potential of hyperspectral imaging and machine learning for precise nutrient management in hydrangea cultivation and contribute to sustainable agricultural practices globally.