Comparison of different methods for estimating nitrogen concentration in flue-cured tobacco leaves based on hyperspectral reflectance

- Hyperspectra was applied to analyze the LNC of tobacco using two-year data.
- Identified the central bands sensitive to tobacco LNC and their optimum combination.
- Established new spectral indices thus greatly facilitate monitoring of tobacco LNC.
- The BP neural network based on the new spectral indices was more accurate and stable.

Leaf nitrogen content (LNC) is an important indicator of tobacco quality and is used in the prediction of tobacco yield. Reflectance experiments for flue-cured tobacco were conducted over 2 consecutive years. Leaf hyperspectral reflectance and nitrogen content data were collected at 15-day intervals from 30 days after transplant until harvest. In this work, we identified the central band that sensitive to tobacco LNC and the optimum combination to establish new spectral indices (SR and NDVI), which were used in linear models of the specific ratio vegetation index (SR), normalized difference vegetation index (NDVI), stepwise multiple linear regression (SMLR), and back-propagation (BP) neural network models as independent variable or input factors. The central bands for the LNC were concentrated in the visible range (450-750 nm) in combination with the shortwave infrared range (1450-2500 nm) range. The optimum band combinations for SR and NDVI were (590 and 1980 nm) and (1970 and 650 nm), respectively. The BP neural network model was the most stable and accurate model (R2 = 0.91, RMSE = 0.09, and K¯=0.00). The SR, NDVI, and SMLR models had R2 values of 0.77, 0.76, and 0.86; RMSE values of 0.26, 0.51, and 0.60, and K¯ values of 0.05, 0.11, and 0.14, respectively. The results indicate the possibility of monitoring LNC by combining remote sensing with predictive models.