Detection and assessment of nitrogen effect on cold tolerance for tea by hyperspectral reflectance with PLSR, PCR, and LM models

Knowledge of nutrients effect on freezing tolerance is vital for protecting tea plants against cold injury (CI). Freezing injury treatments on tea leaves with different low temperature and nitrogen (N) concentration was evaluated by hyperspectral imaging based on the reflectance as potential analytical tool. Subsequently, quantitative evaluation of the CI was improved by comparing analyzed results using partial least squares regression (PLSR), principle component regression (PCR) and linear model (LM) models. Reflectance of the CI was obtained by hyperspectral imaging system in the band between 871 and 1766 nm. The results showed that average reflectance increases with the rise of N concentration. A substantial portion of the leaf from the plant with 100% N dosage had the darkest image and resulted in the highest reflectance because the N weakened the negative effect of freezing stress. Out of the five spectral domains tested, the best predictive accuracy for the CI of the tea leaf was achieved by PCR (R2 = 0.9971, RMSE = 0.0609) in 1410–1766 nm wavelength, followed by LM (R2 = 0.9999, RMSE = 0.0805) in 871–1000 nm. The whole interval had the worse prediction accuracy which could be caused by large variations in the data for a specific treatment and high absorption peak occurring around 1450 nm in the reflectance curve. The averages, R2 and RMSE for all the three statistical models showed that the worse prediction accuracy occurred in 1410–1766 nm, followed by the whole interval 871–1766 nm. The prediction accuracy was low which could be due to the strong water absorption peaks that appeared in both range of wavelengths. The models without absorption bands had improved correlation coefficient and reduced RMSE values between the measured and the predicted CI (R2 = 0.9676; RMSE = 0.3067) compared to the model developed with the reflectance values in the entire waveband (R2 = 0.9504; RMSE = 0.6629). The reflectance values where absorption bands occurred had detrimental effect on the model performance, which resulted in the lower correlation values and larger errors. This study has demonstrated that the wavelength at which absorption bands occur can influence model performance significantly which is a step towards real-time implementation of the technique.