Drone-Based Oil Palm Harvest Estimation Using Enhanced YOLOv8 and Cloud-Integrated Precision Agriculture System

Abstract

Oil palm (Elaeis guineensis) is the dominant source of vegetable oil worldwide, contributing nearly 40% of global production and playing a vital role in economic development and food supply chains. Accurate estimation of fresh fruit bunches (FFB) during harvesting, known as raw fruit census, is critical for optimizing labour allocation, transportation resources, and yield forecasting in commercial plantations. Traditional census methods rely on manual fruit counting, which is labour-intensive, error-prone, and inefficient, particularly for tall trees and large plantation scales. This research introduces an innovative digital framework for harvest estimation by uniting drone vision, state-of-the-art deep learning, and cloud-based decision support. The proposed approach enhances the YOLOv8 object detection framework with GhostNet modules to reduce computational redundancy and a Convolutional Block Attention Module (CBAM) to improve spatial-channel feature extraction under occlusion and lighting variability. The system also integrates DeepSort for multi-object tracking and counting, enabling robust FFB estimation across video sequences. Data collected from oil palm plantations were annotated and pre-processed to train and evaluate the model. Experimental results demonstrate that the proposed YOLOv8-Ghost-CBAM model achieves superior mean Average Precision (mAP50 > 90%) and reduced GFLOPs compared to baseline YOLOv8 and EfficientDet models, while maintaining real-time performance on drone-mounted devices. Furthermore, cloud-based dashboards allow plantation managers to access real-time harvest forecasts and resource planning insights. This research contributes to the advancement of precision agriculture by bridging the gap between detection accuracy, deployment efficiency, and operational scalability, offering a practical solution for sustainable oil palm plantation management.