Spectral and foliar analysis using multiple machine learning classifiers for mature oil palm treated with nitrogen fertilizer
In assessing the leaf biochemical properties, spectral analysis has been explored as the non-invasive alternative to destructive leaf analysis. This study aims to develop spectral-based classification models to estimate oil palm’s chlorophyll (chl) and nutrient status via machine learning (ML). In t...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2021
|
| Subjects: | |
| Online Access: | http://psasir.upm.edu.my/id/eprint/98770/1/FP%202021%2057%20-%20IR.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In assessing the leaf biochemical properties, spectral analysis has been explored as the non-invasive alternative to destructive leaf analysis. This study aims to develop spectral-based classification models to estimate oil palm’s chlorophyll (chl) and nutrient status via machine learning (ML). In this study, different nitrogen (N) rates were applied to 8 and 11 years old. The leaf nitrogen (N), phosphorus, potassium, magnesium, calcium, chl a, chl b, total chl content, and relative chl content were measured from frond 9 and 17. Meanwhile, spectral reflectance in visible (Vis) and near-infrared (NIR) were measured at three spatial scales: leaf (spectroradiometer), canopy (unmanned aerial vehicle (UAV)), and scene (SPOT-6 satellite).
The objectives of this study are to; 1) evaluate the leaf spectral data and ML (Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM)) in classifying N status; 2) examine SPOT-6-derived spectral indices (SIs) in monitoring N using LDA and SVM; 3) discriminate chls status via spectroradiometer and ML (Random Forest (RF) and Decision Tree (DT)); 4) classify nutrients status via integration of spectroradiometer, ML (Logistic Model Tree (LMT) and Naïve Bayes) and imbalance approaches (Synthetic Minority Over-Sampling TEchnique (SMOTE), Adaptive Boosting (AdaBoost) and combination of SMOTE and AdaBoost (SMOTE+AdaBoost)); while 5) assess leaf and canopy spectral models in discerning the chls and nutrients status via LMT-AdaBoost.
In objective 1, there is a clear trade-off between the number of the sensitive-N spectral features with LDA accuracy and the N-sensitive features responses were palm age-dependent. For objective 2, the Vis index (Blue Green Red Index (BGRI1 and BGRI2)) (79.55%) and Vis+NIR index (Normalized Difference Vegetation Index, Normalized Green, Infrared Percentage Vegetation Index, and Green Normalized Difference Vegetation Index) (81.82%)) were the best SIs to assess N status of young- and prime-mature palms, respectively via SVM. The SVM was superior to LDA in categorizing the N status and the N-sensitive SIs tested were palm age-dependent. Results from objective 3 showed that chl-sensitive features are often positioned at the red edge and RF outperformed the DT in discriminating the chl status (96.05 to 98.08%). Meanwhile, the best discrimination of nutrients status (objective 4) was achieved via the LMT-SMOTE+AdaBoost (76.13 to 100.00%). Also, the effect of frond-age was prominent in both chls and nutrients studies via spectroradiometer. The UAV study (objective 5) highlighted the capability of SIs was greater than the raw band in assessing the chls and nutrients status of oil palm (74.64 to 100.00%). In comparing the competency of leaf and canopy spectral models, the former manifested robust accuracies (98.53 to 100.00%), yet, the latter model offered wide coverage with a lesser number of spectral features, elucidating by the maximum difference of 25.36%.
In a nutshell, the leaf scale is portrayed as the best platform in discriminating the chls and nutrients status followed by canopy and scene. The canopy and scene scales could assess the leaf biochemical properties with satisfactory accuracy. It is also suggested that the LMT-SMOTE+AdaBoost was the finest classifier in evaluating the chls and nutrients of oil palm. |
|---|