Oil Palm Tree Recognition and Counting Based on the Aerial View with XAI

Chong, Jonathan Hao Jie (2023) Oil Palm Tree Recognition and Counting Based on the Aerial View with XAI. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Full Text -Jonathan Chong Hao Jie.pdf Restricted to Registered users only Download (3MB)

Abstract

The palm crop plantation plays a significant role in the market for edible oils worldwide. The oil palm density is important data as it yields crucial information on palm productivity, fertilizer requirements, manpower requirements, and requirements for other activities. With the advancement of technology, oil palm tree counting can be completed by counting the oil palm trees in aerial images using a deep learning object detection model. The research on detecting and recognizing both young and mature oil palm trees with a single deep learning mode is still lacking. Besides that, almost all the previous related work for oil palm tree recognition that uses deep learning methods do not have a visual explanation. Therefore, in this research, YOLOv5, YOLOv6, and YOLOv7 are used to recognize and count the number of young and mature oil palm trees with XAI. Among the YOLO models, YOLOv7e6e-RepVGGN has the highest AP and AR. Gradient free CAM algorithms such as Eigen-CAM, Ablation-CAM, and Score-CAM are used to provide a visual explanation of the YOLO models that have been trained. Ablation-CAM generates the highest quality heat maps for all three target layers of YOLOv7 as compared to Eigen-CAM and Score-CAM.