Small target detection algorithm based on SAHI-Improved-YOLOv8 for UAV imagery: A case study of tree pit detection
The application of deep learning in tree pit detection of unmanned aerial vehicle (UAV) images has problems such as dense distribution, high density, small size, false detections, missed detections, and high localization error. To address these challenges, this paper proposes a small target detectio...
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/S2772375525004125 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The application of deep learning in tree pit detection of unmanned aerial vehicle (UAV) images has problems such as dense distribution, high density, small size, false detections, missed detections, and high localization error. To address these challenges, this paper proposes a small target detection algorithm based on SAHI-ImprovedYOLOv8 for detecting tree pits, which includes Slicing Aided Hyper Inference (SAHI), Focal loss, Spatial Pyramid Pooling Concurrent Spatial Pyramid Convolution (SPPCSPC), and Convolutional Block Attention Module (CBAM). The accuracy of identification and positioning can be improved by using the SAHI via cutting high-resolution UAV imagery into slices that match the detection model, avoiding the loss of small target detail caused by direct downsampling. The identification accuracy is improved by using the Focal loss and CBAM, SPPCSPC to mitigate the data imbalance, strengthen key semantic features, and realize fine-grained information enhancement. The experimental results show that the SAHI-Improved-YOLOv8 model outperforms YOLOv3, YOLOv5, YOLOv8, YOLOv10, YOLOv11 and SAHI-YOLOv8 with a Precision of 85.17 %, a Recall of 85.07 %, a AP50–90 of 78.63 % and a F1 score of 85.12 %. In conclusion, the SAHI-Improved-YOLOv8 has the capability of efficiently processing high-resolution images, which alleviates the problems of high density of small targets, false detections, missed detections, and high localization error. In practical applications, the SAHI-Improved-YOLOv8 model performs excellently in tree pit detection in UAV imagery, significantly reducing false detections and missed detections, and providing reliable technology support for large-scale forest management. |
|---|---|
| ISSN: | 2772-3755 |