Oil palm leaf nutrient estimation using optical sensors

• Main Author: Khorramnia, Khosro
• Format: Thesis
• Language: English
• Published: 2014
• Subjects: Oil palm; Oil palm - Nutrition - Research; Optical sensors
• Online Access: http://psasir.upm.edu.my/id/eprint/48052/1/FK%202014%2038.pdf; http://psasir.upm.edu.my/id/eprint/48052/
• Institution: Universiti Putra Malaysia

Leaf sampling and chemical analysis are common practical methods of assessing the nutrient status of an oil palm leaf. The oil palm foliar analysis technique is expensive and does not provide enough information for individual trees and site-specific fertiliser management. We conducted three experiments in the field and the laboratory, using four different optical sensors. Field measurements were performed three different times at the Universiti Putra Malaysia (2.979917°N, 101.7297833°E),at an agricultural park, and at Sime Darby Co. located at (2.8673°N, 101. 3674°E)Carey Island, Malaysia. Specific objectives of this research were to: (i) evaluate the performance of various spectral bands and indices for measuring N, P, K, Mg, Ca and B status in oil palm fronds using four available active sensors (GreenSeekerRT505, SPAD 502 Plus, Multiplex3 and Spectroradiometer FieldSpec3, Hi-Res ASDi) under laboratory conditions; (ii) to compare the performance of developed models using various spectral bands and indices for measuring N, P, K, Mg, Ca and B status in oil palm fronds. Four modeling techniques, partial least square, stepwise multiple linear regression, artificial neural network and linear discriminant analysis,applied to training datasets for leaf N, P, K, Mg, Ca and B prediction analysis; and (iii) assess model performance on test datasets by testing the correlation of four models of predicted nutrient results with measured nutrients. The next step in model assessment is to compare the effectiveness of modeling methods using the receiver operating characteristic (ROC) method for oil palm leaf nutrient predictions. At the first and second measurements, only GreenSeeker and SPAD502 plus were utilized to develop leaf nutrient estimation models, that was not promising. At the third measurement, spectroradiometer and Multiplex3 were added. Spectral data and indices processed and screened using stepwise multiple linear regressions (SMLR). Then, feature datasets were analysed using artificial neural network (ANN). The maximum accuracies of estimations were N=77%, using spectroradiometer and ANN, P =100%, using spectroradiometer and ANN, K=75%, using Multiplex3 and ANN, Mg=77%, using Multiplex3 and ANN, Ca=98%, using Multiplex3 and ANN and B=91%, using spectroradiometer and ANN. The reliability assessment of models using ROCs and according to their AUCs values were N= 0.83, P= 1.0, K= 0.84, Mg= 0.80, Ca= 0.95 and B= 0.95. Linear discriminant analysis (LDA) applied to training datasets of screened spectroradiometer and Multiplex3 by using entire data for discriminant analysis and using stepwise method, to reduce number of independent variables. Among the three different modelling methods, SMLR, neural network and LDA, neural network models gave higher accuracies to estimate leaf nutrient status. In case of designing and fabricating affordable sensors, LDA could be useful method to develop estimation models using indices (as predictors) provided by Multiplex3, for B. Using neural network method the minimum predictors to estimate B status was 13 indices. In comparison with neural network model, LDA method needs only three Multiplex3 indices (YF_R, FRF_R and SFR_R) and NDVI. Correctly classified samples for N, P, K, Mg, Ca and B using LDA were 50% (Multiplex3), 89% (Spectroradiometer), 70%(Multiplex3), 68%(Multiplex3), 75% (Spectroradiometer or Multiplex3) and 80% (Multiplex3) respectively.