Analyzing the performance of statistical models for estimating leaf nitrogen concentration of Phragmites australis based on leaf spectral reflectance

Abstract

Nitrogen is an essential nutrient for plant growth and development. Rapid and nondestructive monitoring of nitrogen nutrition in plants using hyperspectral remote sensing is important for accurate diagnosis and quality evaluation of plant growth status. The sensitive bands of leaf nitrogen concentration varied in different plants. However, most of the current studies are concentrated on crops, and only a few studies focused on wetland plants. This study investigated the accuracy of the most common univariate, stepwise multiple linear regression, and partial least squares regression models for predicting leaf nitrogen content in a wetland plant reed (Phragmites australis) by testing the accuracy of all the models through leave-one-out cross validation coefficient of determination, root mean square error and relative error. It found that: (i) sensitive bands responding to leaf nitrogen concentration were concentrated in the green and red regions of visible light; (ii) for univariate regression models, the quadratic polynomial model based on the difference spectral index composed of 665 nm and 680 nm had the highest predictive accuracy (the validation coefficient of determination was 0.7535); (iii) for multivariate regression models, the stepwise multiple linear regression models had superior predictive accuracy to the partial least squares regression models, and the stepwise multiple linear regression model with first derivative reflectance was optimal for predicting leaf nitrogen concentration (the validation coefficient of determination was 0.7746, the validation root mean square error was 0.2925, and the validation relative error was 0.0804). The findings provide a scientific basis for rapid estimation and monitoring of leaf nitrogen concentration in P. australis in a nondestructive manner.