Publication: Performance analysis of yolo and ssd-based deep learning models for detection of oil palm trees in drone images
Loading...
Date
2025-07-01
Authors
Istiyak Mudassir Shaikh
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This study explores the use of advanced deep learning models for detecting and counting oil palm plants in precision agriculture using drone-based high-resolution images. The motivation stems from the limitations of manual monitoring methods, which are time-consuming, error-prone, and not feasible for large-scale plantations. Given Malaysia’s significant role in global palm oil production, efficient and automated detection systems are essential to support sustainable plantation management. The primary challenge is to accurately identifying oil palm trees in complex conditions, such as overlapping canopies, dense vegetation, varying lighting, and similar surrounding plants. These factors limit traditional image processing techniques, prompting the use of robust deep learning frameworks. This study evaluates four state-of-the-art object detection models: YOLOv5x, YOLOv7, YOLOv8, and SSDv2FPN, selected for their real-time detection capabilities and accuracy in agricultural environments. Two datasets were used: a smaller set of 10 drone images containing 79 annotated palm trees, and a larger dataset of 482 images with 5,233 trees. Evaluation metrics included True Positives, False Positives, False Negatives, Precision, Recall, F1-Score, and Detection Time. SSDv2FPN achieved perfect precision at 100% with an F1-Score of 89.49%, but required 83 seconds per image, which limits its suitability for real-time applications. In contrast, YOLOv5x, YOLOv7x, and YOLOv8x detected palm trees in relatively lower execution time of 16, 12, and 14 seconds respectively, with YOLOv5x achieving an F1-Score of 97.36%. These results demonstrate the clear advantage of YOLO models with regard to high speed execution. On the larger dataset, YOLOv8 models outperformed other frameworks, thereby achieving F1-Scores between 97.36% and 99.31%, precision values ranging from 99.27% to 99.70%, and recall rates between 95.89% and 99.36%. Among the YOLOv8 variants, YOLOv8s and YOLOv8n demonstrated the fastest detection times of 28 and 33 seconds, respectively, effectively balancing rapid inference and detection performance. This makes them ideal for deployment in practical agricultural monitoring systems.