Machine Learning Approach for BSR Disease Classification of Palm Oil Trees Using RGB Image Analysis

Wong, Jian Wai (2023) Machine Learning Approach for BSR Disease Classification of Palm Oil Trees Using RGB Image Analysis. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Full Text-Wong Jian Wai.pdf Restricted to Registered users only Download (9MB)

Abstract

Malaysiawas responsible for 25.8% and 34.3%, of the global production and exports of palm oil in 2020 respectively. However, there are several diseases associated with palm oil that can result in considerable losses of up to US$500 million annually in Southeast Asia. One of the diseases brought on by the fungus Ganoderma boninese is Basal Stem Rot (BSR). Reduced palm growth, pale green foliage, older fronds turning chlorotic, and other symptoms are all signs of BSR. It is crucial to conduct more research on the classification and management of this disease. The nearinfrared spectroscopy approach, terrestrial laser scanning (TLS), and laboratory-based methods are currently employed to classify BSR disease. These techniques are nevertheless expensive, time-consuming, and overly complicated. Therefore, this research aims to examine the potential of RGB images in the classification of BSR disease in palm oil trees by using several traditional machine learning (Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF) and Gaussian Naïve Bayes (GNB). This study made use of the public dataset with a total of 6179 RGB images taken by Unmanned Aerial Vehicles (UAV) that were collected by other researchers. The oil palm trees were divided into four classes by the previous researcher: Dead, Healthy, Small, and Yellow. The problem of the imbalanced dataset was overcome by oversample or augmentation with degrees of 15,30,45,60 and 75 as well as horizontal and vertical flipping. Resizing the image to the scale (100,100), enhancing the brightness and contrast, enhancing the edges, and edge detection for edge-based segmentation are all included in the image preprocessing. Three features are extracted from the RGB images: RGB colour features ( the pixels value of R, G and B channel), Texture features (Using Gray level co-occurrence matrices to obtain the correlation, dissimilarity, homogeneity and energy) and Edge features (Canny Edge Detection). To lower the dimensionality or minimise the features in the dataset while preserving as much of the data as feasible, principal component analysis (PCA) is used. The result showed that the combination of Edge and Texture features as the input to the SVM classifier had the best performance for the BSR disease classification in oil palm trees. The SVM model achieved 97.84 macro average precision, 97.81% of macro average recall, 97.82% of macro average f1-score and 97.88% of overall accuracy. A user-friendly and responsive local web application is developed in this study using Tkinter to boost the machine-learning model’s acceptance throughout the oil palm sector. The research’s findings provide a potentially cost-effective and efficient method for BSR disease classification in palm oil trees, aiding in disease management and prevention.