Machine-learning approach using thermal and synthetic aperture radar data for classification of oil palm trees with basal stem rot disease

• Main Author: Che Hashim, Izrahayu
• Format: Thesis
• Language: English
• Published: 2021
• Subjects: Oil palm - Technological innovations; Machine learning; Spatial data mining
• Online Access: http://psasir.upm.edu.my/id/eprint/103998/1/IZRAHAYU%20BINTI%20CHE%20HASHIM%20-%20IR.pdf; http://psasir.upm.edu.my/id/eprint/103998/

The fast growth of oil palm has resulted in its development as a strategic global commodity. Oil palm creates export revenues and strengthens the economies of numerous nations, especially Indonesia and Malaysia. However, oil palms are susceptible to basal stem rot (BSR) caused by Ganoderma boninense (G. boninense), the most dangerous oil palm disease. This disease has been a cause for concern as it has caused significant tree mortality in several plantations in Malaysia. Given that there is currently no effective cure for this disease, the only viable solution is to prolong the life of oil palm trees. This study explored the early detection of the BSR using thermal images and an ALOS PALSAR-2 image with dual-polarization, Horizontal transmit and Vertical receive (HV), and Horizontal transmit and Horizontal receive (HH). The study was conducted in Seberang Perak, part of Felcra Seberang Perak 10, and is located in Perak, Malaysia. Initially, an experiment was carried out to (1) identify the potential temperature variables; (2) identify the potential backscatter variables; (3) utilize the imbalance data approach like Random under-sampling (RUS), Random oversampling (ROS), Synthetic Minority Oversampling (SMOTE) and AdaBoost; and (4) evaluate the performance of machine learning (ML) classifiers Naïve Bayes (NB), Multilayer Perceptron (MLP), as well as Random Forest (RF) in classifying the stages and severity levels of G. boninense. The sample size was comprised of 55 non-infected trees and 37 infected trees. During the field experiments, oil palm tree samples of non-infected (T0), mild infected (T1), moderate infected (T2), and severe infected (T3) were measured using the FLIR T620 IR infrared thermal imaging camera to obtain the temperature of the oil palm trees. The temperature variation for each thermal image was examined using FLIR ResearchIR Max, the camera manufacturer's software, and feature extraction for each thermal image was extracted using FLIR Tools in the FLIR ResearcherIR environment software. The backscattering value of each tree was then extracted from the ALOS PALSAR-2 image. Using the Extract Multi Values tool in ArcGIS, the backscattering value for each oil palm point was derived from the processed ALOS PALSAR-2 image. As the ALOS PALSAR-2 image was evaluated with dual-polarization (HH and HV), each digitized point has two distinct backscatter data with four severity levels (T0 to T3). The machine learning algorithm consistently performs well when presented with a well-balanced dataset. In an imbalanced dataset, one of the two classes contains fewer total samples than the other class. The sampling-based method, also known as the data level method, is used to deal with this problem. In this study, the resampling method and ensemble procedure relied entirely on the Waikato Environment for Knowledge Analysis (WEKA) version 3.8.5 software. The classification is performed using the derived features from the thermal images and the backscatter features. The extracted features serve as predictors and the status of oil palm as a response. To identify non-infected and BSR-infected trees, the WEKA tool version 3.8.5 was used for classification. The classifiers evaluated in this study were Nave Bayes (NB), Multilayer Perceptron (MLP), and Random Forest (RF). Two datasets, for training and testing, were both classified. We divided the dataset into a training dataset of 70% and a test dataset of 30%. The classification was done with 10-fold cross-validation to avoid overfitting and get unbiased prediction error estimates. This was the recommended validation method for the small dataset. This study, therefore, detailed the description of the confusion matrix as an alternative in terms of the rate of success of the non-infected and BSR-infected tree together with the balanced classification rate (BCR) or balanced accuracy, the precision-recall curve (PRC), and receiver operating characteristics (ROC) curve region (AUC) to evaluate different classifier and imbalanced approaches and measure their performance. The study found that the Tmax, Tmin feature is the most beneficial concerning other temperature characteristics for classifying non-infected or infected BSR trees. In the meantime, the HV feature is most advantageous for classifying non-infected or infected BSR trees compared to other backscatters. Compared with a single approach and other approximate imbalance data approaches, the ROS approach improves BCR, AUC, and PRC data results in datasets. Next, all classifier models were employed in classifying the BSR disease severity using the combination of the best features of temperature (Tmax, Tmin), backscatter features (HV), and significant ground-based data (DbH and soil moisture) with a single and ROS approach. In conclusion, all three ML methods can classify the oil palm with severe BSR disease with an outstanding result using the ROS approach. Meanwhile, the MLP was found to be the ideal model with a BCR value of 0.964, AUC and PRC having the same value of 1.000, model accuracy of 96.43%, and a Kappa coefficient of 0.95. The MLP classifier model also had a high success rate, whereby it correctly classified 85.71% (T0-healthy), 100% (T1-mild infected), 100% (T2-moderate infected), and 100% (T3-severe infected). This study concluded that for the early detection of BSR, a significant degree of accuracy was obtained. Infected palms are asymptomatic throughout the disease's early stages, making disease detection challenging. The survival of affected trees must detect BSR at the mild infected (T1) stage. A meaningful conclusion of this study is that the ROS technique can differentiate the severity of mild infection (T1) compared to a single approach that is incapable of doing so. The main benefit of this study is the development of an appropriate model for early identification and severity classification of BSR disease in oil palms via remote sensing and data mining approaches rapidly and cost-effectively.