Dissection of hyperspectral reflectance to estimate nitrogen and chlorophyll contents in tea leaves based on machine learning algorithms

Abstract

Nondestructive techniques for estimating nitrogen (N) status are essential tools for optimizing N fertilization input and reducing the environmental impact of agricultural N management, especially in green tea cultivation, which is notably problematic. Previously, hyperspectral indices for chlorophyll (Chl) estimation, namely a green peak and red edge in the visible region, have been identified and used for N estimation because leaf N content closely related to Chl content in green leaves. Herein, datasets of N and Chl contents, and visible and near-infrared hyperspectral reflectance, derived from green leaves under various N nutrient conditions and albino yellow leaves were obtained. A regression model was then constructed using several machine learning algorithms and preprocessing techniques. Machine learning algorithms achieved high-performance models for N and Chl content, ensuring an accuracy threshold of 1.4 or 2.0 based on the ratio of performance to deviation values. Data-based sensitivity analysis through integration of the green and yellow leaves datasets identified clear differences in reflectance to estimate N and Chl contents, especially at 1325-1575 nm, suggesting an N content-specific region. These findings will enable the nondestructive estimation of leaf N content in tea plants and contribute advanced indices for nondestructive tracking of N status in crops.