Deep Learning Approach for BSR Disease Detection of Oil Palm Tree Using RGB Image Analysis

Ong, Win Kent (2023) Deep Learning Approach for BSR Disease Detection of Oil Palm Tree Using RGB Image Analysis. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Full Text-OngWinKent.pdf Restricted to Registered users only Download (16MB)

Abstract

The oil palm tree is one of the most valuable and highly important economic plants that act as the main export of the country, especially the South East Asia countries such as Malaysia and Indonesia. The palm oil produced from the oil palm tree also plays the important role in the daily life product manufacturing process. Hence the production of the oil palm tree is very crucial. However, the BSR disease has been identified as the main threat to the oil palm tree causing the infected trees to reduce the production of oil palm trees finally leading to death. If treatment has not been applied at its earlier stage of infection. In doing so a high accuracy and efficiency of the oil palm tree BSR disease detection solution needs to be developed. The proposed oil palm tree BSR disease detection algorithm implements the deep learning-based segmentation algorithm as the backbone to segment the input image in form of a probability map. Afterwards, the segmented image was further postprocessing through the proposed image postprocessing technique to finalise the detection output. A novel deep learning-based segmentation architecture namely Multi- convolution Residual U-Net(M-CR U-Net) has been proposed in this study. The modification focus on the convolution block by implementing multi-convolution kernel sizes with a skip connection to solve the main issue existing in the original U-Net architecture. The segmentation performance of the proposed M-CR U-Net is surpass several U-Net family architectures in oil palm tree BSR disease detection has been proven in this study. Moreover, the proposed M-CR U-Net was able to achieve a modest improvement in oil palm tree BSR disease detection after applying the image postprocessing algorithm, especially in the oil palm tree’s extremely crowded region. A novel image postprocessing algorithm to separate the overlapping region between segmented regions has been proposed in this study. The performance of the proposed method in this study was compared against the state of art algorithm which is Faster-RCNN implemented in the previous study by using the same datasets. The proposed method in this study yields a higher performance compared to the reference method in oil palm tree BSR disease detection. Even using class-imbalanced datasets. The proposed method in this study shows high potential in oil palm tree BSR disease detection through UAV images which could improve oil palm tree plantation management significantly.