Non-destructive hyperspectral measurement model of chlorophyll content for citrus leaves

Abstract: Traditional methods of obtaining chlorophyll content of citrus leaves require grinding citrus leaves and applying chemical titrations, which would be harmful to citrus trees and time-consuming. Besides, it's difficult to integrate those chemical methods into variable spraying system as a feedback subsystem. In this paper, we discuss several rapid and non-destructive methods in obtaining chlorophyll content of citrus leaves by using hyperspectral analysis system. Hyperspectral technology obtains synchronously spectrum in continuous space, where we can derive crop growth information visually in a non-destructive way. In this paper, the modeling of chlorophyll content of citrus leaves based on the hyperspectrum was discussed. Field experiments were conducted on 117 planted Luogang citrus trees in the Crab Village of Luogang District, Guangzhou City, Guangdong Province. Hyperspectral reflectance and chlorophyll content of citrus leaves were measured by spectrometer (ASD FieldSpec 3) and traditional spectrophotometry, respectively, during four different growth periods corresponding to germination period, stability period, bloom period and harvesting period. In this way, each sample was presented as an instance-labeled pair, where a high-dimensional vector was regarded as the descriptor along with the measured value of chlorophyll content. All the collected samples constituted a large-scale dataset with totally 468 tuples, 80% of which were utilized as the training set and remaining 20% as the testing set. The model constructed relied on the training set and the testing set was evaluated respectively. Using original spectrum and its transformations as input vector, two models, support vector regression (SVR) based on principle component analysis (PCA) and partial least square regression (PLSR) based on the wavelet denoising were adopted, where PCA was adopted for dimension reduction and the wavelet denoising technique removed high-frequency noise. The two models (SVR and PLSR) were then applied to the final regression analysis for predicting chlorophyll content. The best coefficient of determination (R2) of the calibration set and a validation set of the entire growth period were up to 0.8713 and 0.8670, the root-mean-square error (RMSE) was 0.1517 and 0.1544 respectively. Some main conclusions were obtained: first, when the original reflectance spectrum was used as the input vector and the energy ratio remained 96% for PCA in germination period and stability period, 99% for PCA in bloom period, harvesting period and the whole growth period, SVR with the radial basis function (RBF) as the kernel function achieved the best performance. Second, the wavelet denoising for hyperspectrum data could improve the model performance to some extent. When "sym8" was used as the wavelet basis function, "rigrsure" as the threshold selection, "sln" for rescaling using a single estimation of level noise based on first-level coefficients as the threshold rescaling project and the decomposition layer was 5, PLSR achieved the best result in this research and the coefficient of determination of calibration set and the validation set of the whole growth period were up to 0.8706 and 0.8531, which increased by 8.3% and 9.3% compared with the model without the wavelet denoising. Third, comparative tests between our best model and other models demonstrate the validity and robustness of the two models we derived. Further experimental results revealed that these two models were superior to principle component regression (PCR), stepwise multiple linear regression (SMLR) and back propagation (BP) neural networks. Finally, hyperspectral technology could obtain accurate chlorophyll content of citrus leaves rapidly, quantitatively and non-destructively, our research may provide a theoretical basis for nutrition surveillance of citrus growth.