Harnessing YOLO for Loose Fruits Detection: Boosting Productivity in Palm Oil Plantations

Abstract

This study explores the use of the YOLOv5 algorithm to detect loose fruits palm oil in plantations, improve collection efficiency, and reduce reliance on manual labor. Loose fruits have high economic value due to their oil content, but they are often not collected, resulting in financial losses and the growth of weeds around oil palm trees. Conventional manual methods are often ineffective; thus, automatic detection is required. The dataset comprises 525 images of loose fruits, covering complex conditions on plantations, and secondary data from Roboflow, including nontarget objects, such as rocks and dry leaves, to train the model to accurately distinguish loose fruits. During training, several parameters were optimized: batch size, learning rate, optimizer type, weight decay, and momentum, which showed that small batch sizes provided higher mAP50 values, whereas large batches increased mAP50:95, indicating the ability to detect objects at higher levels of complexity. Testing in a scenario with loose fruits clearly visible and partially covered with leaves or grass showed that YOLOv5 could detect loose fruits well despite some constraints in the obstructed environment. The results of this study show that, with appropriate parameter settings, YOLOv5 can be a reliable basis for automated loose fruit collection systems, potentially improving operational efficiency in oil palm plantations and contributing to the development of automation in the agricultural sector.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint