Estimation of chlorophyll content in apple leaves base on spectral feature parameters

Abstract: Chlorophyll content is an important parameter for evaluating the growth status using spectral reflectance feature. The rapid, non-destructive and accurate monitoring of chlorophyll content using hyperspectral reflectance has become an important research content for monitoring the growth of fruit trees. The object of this study was to analyze the relevance of chlorophyll content and the original spectrum of apple leaves and its transformation forms and to select optimum spectral parameters. Chlorophyll content model was built and verified by random forest (RF), partial least square (PLS), back propagation (BP) neural network and support vector machine (SVM). Parameters of samples including spectral reflectance of leaves and the concurrent apple leaves chlorophyll content were acquired in Tai'an, Shandong, China during apple growth seasons in 2012 and 2013. The result showed: 1) The optimum spectral parameters between chlorophyll content and the original spectrum reflectance (R) of apple leaves were 554 and 708 nm, and the correlation coefficients of that were ?0.46 and ?0.66 respectively. The optimum spectral parameters between the chlorophyll content and the logarithm of reciprocal of spectra of apple leaves were 554 and 708 nm, and the correlation coefficients of that were 0.46 and 0.66 respectively. The optimum spectral parameters between chlorophyll content and the first order differential (D) reflectance spectra of apple leaves were 535 (trough), 569 (peak), 700 (trough) and 749 nm (peak), and the correlation coefficients of that were ?0.66, 0.64, ?0.69 and 0.76 respectively. The optimum spectral parameters between chlorophyll content and the continuum removal (CR) reflectance spectra of apple leaves were 557 (trough) and 708 nm (trough), and the correlation coefficients of that were ?0.35 and ?0.73, respectively. 2) The out-of-bag importance between chlorophyll content and reflectance spectra was analyzed using out-of-bag data of RF, the size order of out-of-bag data was D749 > CR708 > D569 > D700>D535 > CR557 > log(1/708) > log(1/554) > R554 > R708, the maximum and minimum were D749 and R708, respectively, and the corresponding values were 166.28 and 7.34, respectively. Based on out-of-bag data analysis, the D749, CR708, D569, D700 and D535 were chosen to build chlorophyll content estimation model using RF, PLS, BP, and SVM. The result showed that the R2, RMSE (root mean square error) and RE (relative error) were 0.94, 0.34 mg/dm2 and 0.08% respectively for RF-estimation model; the R2, RMSE and RE were 0.61, 0.78 mg/dm2 and 0 respectively for PLS-estimation model; the R2, RMSE and RE were 0.66, 0.75 mg/dm2 and 0.25% respectively according to BP-estimation model; the R2, RMSE and RE were 0.60, 0.81 mg/dm2 and 0.70% respectively according to SVM-estimation model. The accuracies of RF, PLS, BP and SVM validation model were compared. The R2 of RF, PLS, BP and SVM model was 0.86, 0.91, 0.60 and 0.66, respectively; the RMSE of RF, PLS, BP and SVM model was 0.79, 0.75, 1.18 and 1.20 mg/dm2, respectively; the RE of RF, PLS, BP and SVM model was 1.31%, 6.68%, 3.19% and 0.46%, respectively. The study showed that the accuracy of RF estimation model is much higher than PLS, BP and SVM. The stability of the RF validation model is also higher than that of the PLS and BP validation model, which is close to the PLS regression. Overall, the RF algorithm has better performance than PLS, BP and SVM algorithm. Therefore, using hyperspectral technology with RF algorithm can estimate apple leaf chlorophyll content more rapidly and accurately and provide a theoretical basis for rapid nutrition diagnosis and growth monitoring.