Detection Model of Nitrogen Content in Maize Leaves Based on Hyperspectral Imaging

Lifeng Wang,Changli Zhang,Yue Zhao,Yuzhu Song,Runtao Wang,Zhongbin Su,Shuwen Wang

DOI: https://doi.org/10.13427/j.cnki.njyi.2017.11.027

2017-01-01

Abstract:Visible and near infrared (Vis-NIR) hyperspectral imaging system was carried out to rapidly detect the nitrogen(N) content in maize leaves.The data of average spectral reflectance, which were acquired by analyzing hyperspectral information(400~1000nm wavelength) of 240 leaf samples within region of interest(ROI), were preprocessed by three different methods, including First derivative (1st Der),Second derivative(2nd Der) and standard normal variate (SNV), to improve the signal to noise ratio.Several estimation models have been built based on different parameters using different algorithms for comparison, including: (1) Using three amplitude parameters (Dλr、Dλy、Dλb) to establish a single variable regression model and multivariate linear regression model (2) a nitrogen estimation model based on partial least squares (PLS) in the whole wavelength region of 400 ~ 1000nm;(3) Among 203 wavelengths, only 10 wavelengths were selected by successive projections algorithm (SPA) as the effective wavelengths for N prediction.Based on these effective wavelengths, partial least squares (PLS) calibration model was established for the determination of N content.;(4) Using principal component analysis (PCA) to extract the 11 main components which were applied to establish 1st Der-PCA-PLS model.A nitrogen estimation model based on PCA with PLS has the lowest prediction accuracy and prediction results are not feasible.The results indicate that the optimal prediction performance is achieved by PLS model in the whole wavelength region of 400 ~ 1000nm, which has a correlation coefficient of 0.967 and a root mean squares error of predicted of 0.024.The SPA-PLS model can provide a very close prediction result while the calibration computation has been significantly reduced and the calibration speed has been accelerated sharply.This study provibes valuable information for rapid and non-destructive nitrogen detection in crops.For simple experience models based on amplitude parameters ,although there is a slight gap between the prediction performance and that of the PLS model in the whole wavelength region of 400~1000nm, they also have their own unique advantages which should be thought highly of: these models are much simpler and thus the calibration computation is reduced significantly, they can perform an important function under circumstances in which increasing modeling speed and reducing calibration computation operand are more important than improving the prediction accuracy, such as the development of portable devices.