Classifying the Health Status of Oil Palm Trees Using Deep Neural Network

Wee, Marcos Ming Wei (2024) Classifying the Health Status of Oil Palm Trees Using Deep Neural Network. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text MARCOS WEE MING WEI_Full Text.pdf Restricted to Registered users only Download (8MB)

Abstract

Early detection of oil palm tree diseases is crucial to prevent significant losses in Malaysia’s economic income and prevent destructive diseases, such as Ganoderma Basal Stem Rot (BSR), from spreading to other oil palm trees and allows for appropriate treatment and strategized for the disease control accordingly. To conduct high-speed health status inspection, an unmanned aerial vehicle (UAV) is paired with an AI technique known as deep learning to reduce operating cost and time while covering a more expansive region of oil palm trees than the tedious and time-consuming inspection by the naked eye of human experts, especially in a large-scale oil palm plantation. This paper examined the performance of deep learning-based CNN models of VGG16, ResNet50, MobileNetV1, and MobileNetV2 to classify the health status of oil palm trees. As an industrial project, the data acquisition process was done by the industry using a UAV drone with an RGB camera to obtain top-viewed aerial images. There are a total of 2585 dataset images, of which 1202 are labeled healthy, and the other 1383 are labeled unhealthy by human experts from the industry. The dataset was split into a 7:2:1 ratio of training-validation-testing datasets and resized to 224 x 224 pixels. The deep learning (DL) model trained and learned data pattern through the training dataset. The validation dataset was used to fine-tune the hyperparameters and performance evaluation of the trained DL models in an iterative process. The test dataset is a separate, unseen dataset to provide an unbiased final evaluation of the model. The models were compared using performance evaluation criteria such as accuracy, precision, recall, and F1-score to select the optimal model. Among the four models evaluated, MobileNetV1 achieved the best accuracy of 1. Thereafter, based on the MobileNetV1 architecture, the Slender MobileNet with a leaner structure was developed. The model parameters decreased from 2,260,546 to 553,506 compared to MobileNetV1, it boasts a 65.8% reduction in training time, 41.5% reduction in testing time, and still achieved a test result close to MobileNetV1. With an accuracy of 99.61%, precision of 99.28%, recall of 100%, and F1-score of 99.64%, this model has the potential to be employed in a large scale tree health classification system for its high efficiency.