Estimation of foliar nitrogen of rubber trees using hyperspectral reflectance with feature bands

Abstract

Listen

Translate article

Leaf nitrogen is an indispensable parameter for simulating biogeochemical cycling in ecosystems. However, detailed information about leaf nitrogen of rubber trees (Hevea brasiliensis) in the tropical regions is rare. Aims of the current study were to determine the feature bands of foliar nitrogen of rubber trees and to examine the ability of these bands to estimate leaf nitrogen concentration (LNC). In this work, a hybrid variable selection method namely competitive adaptive reweighted sampling (CARS) combined with successive projections algorithm (SPA) (CARS-SPA) was used to extract feature bands for predicting LNC of rubber trees. Two hundred leaf samples were gathered from five fields through April to November (tapping season) in 2014. These samples were divided into calibration dataset (n = 140) and prediction dataset (n = 60) using the Kennard-Stone algorithm. Sixty bands were determined from the first derivative spectrum using the CARS-SPA. Among these bands, not only the commonly used absorption features of chlorophyll, protein, and the red edge position were included, but also those of starch, cellulose, and lignin were selected. All the 60 bands were used as input variables for the partial least squares regression (PLSR) and artificial neural networks (ANN) models to estimate LNC of rubber trees. The determination coefficient of calibration (rc2) and prediction (rp2), and root mean square error of calibration (RMSEC) and prediction (RMSEP), as well as normalized RMSEC (nRMSEC) and normalized RMSEP (nRMSEP) were employed to evaluate performances of these models. Results indicated that CARS-SPA-ANN (rc2 = 0.82, RMSEC = 0.24%, nRMSEC = 7.76%; rp2 = 0.78, RMSEP = 0.22%, nRMSEP = 6.73%) outperformed the other selected models except CARS-PLSR (rc2 = 0.91, RMSEC = 0.17%, nRMSEC = 5.46%; rp2 = 0.80, RMSEP = 0.21%, nRMSECP = 6.44%). However, CARS-SPA-ANN contained much less bands and was more stable than CARS-PLSR. In conclusion, hyperspectral feature bands in combination with ANN could accurately and robustly estimate LNC of rubber trees.