AdaBoost-based Back Propagation Neural Networks Improve the Estimation Accuracy of Pear Leaf Nitrogen Concentration by the In-field VIS-NIR Spectroscopy

Abstract:

For the widespread use of non-destructive determination of leaf nitrogen (N) concentration by the visible near infrared spectroscopy, a high accuracy and robustness of modelling method is urgently needed for dealing with the mixed cultivars in the most pear orchards. In this study, we proposed a new technique to create a model based on the application of Adaboost initiated with classification modelling methods of support vector machine (SVM) and back propagation neural networks (BPNN). The mixed algorithm was expected to be completed by introducing the Adaboost in the procedure of weight distribution over the training examples. The performance was evaluated for estimating leaf N concentration (a total of 1285 samples from different cultivars, growth regions and tree ages) and compared with traditional techniques including vegetation induces (DVI, RVI and NDVI), partial least squares regression, singular SVM and BPNN. Our results demonstrated that the specific absorption valleys and reflection peaks of leaf reflectance were more sensitive to the cultivars than that of different growth regions and tree ages. Moreover, among the eight modelling methods, the Adaboost-BPNN performed the best accuracy in both calibration (R²=0.96, MRE=2.1%) and validation sets (R²=0.92, MRE=3.7%) as well as a steady robustness in the 20 times repeated experiments. In conclusion, our results demonstrate that Adaboost-BPNN can be applied across variables of cultivars, plantation regions and tree ages without the necessity for extensive calibration. This provides us a new insight for dealing with mixed spectral data and opens new possibilities for the nitrogen status assessment of pear trees to better nitrogen management in heterogeneous pear orchards.