Classifying the Health Status of Oil Palm Tress Using AI Techniques

Wong, Joshua Xia Xiong (2024) Classifying the Health Status of Oil Palm Tress Using AI Techniques. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text JOSHUA WONG XIA XIONG_Full Text.pdf Restricted to Registered users only Download (20MB)

Abstract

Oil palm trees, known for their fruit bunches that produce palm oil and palm kernel oil, are extensively cultivated tropical palms. Monitoring the health of oil palm trees relies on manual inspections and is time-consuming. Therefore, Convolutional Neural Networks (CNNs) for image classification are implemented where they could provide a more efficient method for monitoring the health status of oil palm trees. This research identifies VGG16, ResNet50, DenseNet201, and Inception V3 as the models for classifying the health status of oil palm trees. Three datasets were obtained from Aonic Sdn Bhd. The classification dataset comprises 2002 images, with 1001 healthy and 1001 unhealthy oil palm trees respectively. The impact of diversity problem study dataset consists of 582 images, with 200 healthy and 382 unhealthy oil palm trees. The final dataset will be a combination of the classification and impact study of diversity problem datasets. Each CNN model will use the same set of selected hyperparameters, including the number of epochs, learning rate, batch size, and optimizer types. These hyperparameters will be adjusted across 48 different combinations for each CNN model to determine the best-performing set. The best hyperparameter set for each CNN model is determined to be VGG16 with the M21 set, ResNet50 with the M17 set, DenseNet201 with the M45 set, and Inception V3 with the M36 set. The performance of the classification CNN models will be evaluated using metrics such as accuracy, precision, recall, F1-score, and computational time to provide a comprehensive assessment of their effectiveness in distinguishing healthy and unhealthy oil palm trees. Among these CNN models, VGG16 with the M21 set achieves the highest performance results with 0.99 accuracy, a computational time of 5 minutes and 17 seconds, and precision, recall, and F1-score scores of 0.99. The selected models are then tested with a new dataset to study the impact of a dataset diverse problem on the models. The Overall performance of all selected CNN models with the new dataset is considerably lower compared to using only the original dataset. The diverse problem is then addressed by combining the original dataset with the new dataset, resulting in a significant improvement in the overall performance of all CNN models compared to scenario 1.