Prediction of SPAD value in oilseed rape leaves using hyperspectral imaging technique]

Xi-bin Ding,Fei Liu,Chu Zhang,Yong He

DOI: https://doi.org/10.3964/j.issn.1000-0593(2015)02-0486-06

2015-01-01

Spectroscopy and spectral analysis

Abstract:In the present work, prediction models of SPAD value (Soil and Plant Analyzer Development, often used as a parameter to indicate chlorophyll content) in oilseed rape leaves were successfully built using hyperspectral imaging technique. The hyperspectral images of 160 oilseed rape leaf samples in the spectral range of 380 similar to 1030 nm were acquired. Average spectrum was extracted from the region of interest(ROI) of each sample. We chose spectral data in the spectral range of 500 similar to 900 um for analysis. Using Monte Carlo partial least squares(MC-PLS) algorithm, 13 samples were identified as outliers and eliminated. Based on the spectral information and measured SPAD values of the rest 147 samples, several estimation models have been built based on different parameters using different algorithms for comparison, including: (1) a SPAD value estimation model based on partial least squares(PLS) in the whole wavelength region of 500 similar to 900 nm; (2) a SPAD value estimation model based on successive projections algorithmcombined with PLS(SPA-PLS); (3) 4 kind of simple experience SPAD value estimation models in which red edge position was used as an argument; (4) 4 kind of simple experience SPAD value estimation models in which three vegetation indexes R-710/R-760, (R-750-R-705) / (R-750 - R-705) and R-860 / (R-550 x R-708), which all have been proved to have a good relevance with chlorophyll content, were used as an argument respectively; (5) a SPAD value estimation model based on PLS using the 3 vegetation indexes mentioned above. The results indicate that the optimal prediction performance is achieved by PLS model in the whole wavelength region of 500 similar to 900 nrn, which has a correlation coefficient(r(p)) of 0. 833 9 and a root mean squares error of predicted(RMSEP) of 1. 52. The SPA-PLS model can provide avery close prediction result while the calibration computation has been significantly reduced and the calibration speed has been accelerated sharply. For simple experience models based on red edge parameters and vegetation indexes, although there is a slight gap between theprediction performance and that of the PLS model in the whole wavelength region of 500 similar to 900 nm, 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.