Enhanced multi-scale detection of olive tree crowns in UAV orthophotos using a deep learning architecture
Object detection in agriculture is vital for identifying and mapping agricultural areas, especially with the growth of precision farming technologies. Traditional methods for counting and yield estimation are time-consuming and demand significant physical effort, presenting substantial challenges fo...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-12-01
|
| Series: | Smart Agricultural Technology |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2772375525003582 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Object detection in agriculture is vital for identifying and mapping agricultural areas, especially with the growth of precision farming technologies. Traditional methods for counting and yield estimation are time-consuming and demand significant physical effort, presenting substantial challenges for farmers. The use of drones and artificial intelligence, particularly deep learning, has transformed agricultural monitoring, enabling more accurate and rapid analyses. In this study, we introduce an advanced method for detecting tree crowns, focusing on olive trees in farm environments. Our approach is based on an innovative architecture that incorporates a Cross Stage Partial Network (CSPNet) combined with a Feature Pyramid Network (FPN) and Path Aggregation Network (PAN), augmented by DropBlock regularization. Our architecture is tailored for multi-scale object detection from UAV-captured imagery, addressing issues such as small object detection, complex backgrounds, object rotation, scale variations, and category imbalances in both simple imagery and high-resolution orthophotos. These orthophotos are produced by stitching together multiple high-quality images we captured from various angles and altitudes to create a comprehensive and detailed view of the orchard. Our methodology includes splitting images into different sizes (1 × 1, 3 × 3, 6 × 6, and 9 × 9) to enhance analysis and improve detection performance at various scales. This comprehensive approach has enabled us to conduct an in-depth analysis of olive trees, classified into small, medium, and large sizes. The results demonstrate the robustness of our method in addressing common object detection challenges in agricultural contexts, achieving a precision of 92.47 %, recall of 91.40 %, F1-score of 91.93 %, mAP@0.5 of 94.00 %, and mAP@[0.5:0.95] of 87.00 %. These results confirm its reliability for optimizing precision farming practices, including crop condition monitoring and resource management. |
|---|---|
| ISSN: | 2772-3755 |