DETERMINATION OF SIGNIFICANT WAVELENGTHS AND PREDICTION OF NITROGEN CONTENT FOR CITRUS

Abstract

This research was conducted as a preliminary step toward developing a real-time spectral-based nitrogen sensor for citrus trees. Diffuse reflectance of leaf samples, with five nitrogen application rates (0, 112, 168, 224, and 280 kg ha-1), was measured from 400 to 2500 nm using a spectrophotometer in a laboratory environment. A correlation coefficient spectrum, a stepwise multiple linear regression (SMLR) procedure, and the B-matrix in partial least squares (PLS) regression were used to determine important wavelengths. Some wavelengths (448, 669, 719, 1377, 1773, and 2231 nm) were identified by both SMLR and PLS as significant wavelengths for nitrogen detection. The results from the calibration models built by SMLR and PLS showed strong relationships between predicted and actual nitrogen concentration. SMLR performed better on a 20 nm averaged dataset with lower collinearity. With the best calibration model built on wavelengths selected by SMLR from this averaged data set, R2 for the validation data set was 0.839, and the root mean square difference (RMSD) was 0.122%. PLS was more suitable for full spectrum analysis due to its ability to reduce collinearity in the dataset. The calibration model built by PLS on the full spectral region had an R2 of 0.828 and an RMSD of 0.122% for the validation data set.