Deep learning and UAV-Based image recognition for identification of medicinal plants in Gentiana Sect. Cruciata

Rong Ding, Jiangkai Yang, Tianyi Wang, Chenghui Wang, Xi Huang, Shihong Zhong, Rui Gu,
Deep learning and UAV-Based image recognition for identification of medicinal plants in Gentiana Sect. Cruciata,
Computers and Electronics in Agriculture,
Volume 239, Part C,
2025,
111076,
ISSN 0168-1699,
https://doi.org/10.1016/j.compag.2025.111076.
(https://www.sciencedirect.com/science/article/pii/S0168169925011822)
Abstract: Traditional methods for identifying medicinal plant species, such as spectroscopy and chromatography, are often labor-intensive, require specific experimental conditions, and may disturb the natural ecological environment. To address these limitations, this study proposed a non-destructive, UAV-based deep learning approach for large-scale resource assessment of four Gentianaceae species. The YOLO series models, known for their high detection speed and accuracy, were applied for species identification. We evaluated the impact of input image resolution, network architecture, and data augmentation strategies on model performance. The results showed that using 640 × 640 pixel images significantly improved detection accuracy compared to 160 × 160 pixels. The YOLOv5s model achieved the best performance, with a precision of 0.844, recall of 0.801, and mean average precision (mAP0.5) of 0.889 at a resolution of 640 × 640. In contrast, when the input was 160 × 160, the model’s performance declined (precision = 0.76, recall = 0.60, mAP = 0.684), though training time decreased to 0.442 h. Among the improved architectures, YOLOv5s-ShuffleV2 achieved relatively high accuracy (precision = 0.785, mAP0.5 = 0.764) with fewer parameters (3.19 million), offering a lightweight solution for real-time applications. The YOLOv5s model remained the fastest and most accurate model overall (mAP = 0.889, training time = 0.629 h). Data augmentation further improved model generalization across environmental conditions. Applying the optimized model for resource assessment in 40 regions, we achieved an overall mAP0.5 of 0.798 and accuracy of 0.901, with an R2 of 0.98. Among the four target species: Gentiana straminea Maxim. (GsM), Gentiana crassicaulis Duthie ex Burkill. (GcDB), Gentiana siphonantha Maxim. ex Kusn. (GsMK), and Gentiana officinalis Harry Sm (GoHM), GsM achieved highest detection accuracy (mAP0.5 = 0.866), while GoHM was the most challenging (mAP0.5 = 0.742, recall = 0.557). This approach demonstrates the potential for large-scale, non-destructive surveys of wild Gentianaceae resources and offers significant value for similar medicinal plant resource assessments.
Keywords: YOLO; Gentianaceae; Medicinal plant identification; Large-scale plant survey; Data augmentation