Detecting Plant Species in the Field With Deep Learning and Drone Technology – Plateau

Back to the Promont Interreg IPA Adrion programme website

Interreg Banner

Promont Plateau

Protection and RegeneratiOn of MOuNTains

Supporting a greener and climate resilient Adriatic-Ionian region

Back to Good Practices

Detecting Plant Species in the Field With Deep Learning and Drone Technology

Title	Detecting Plant Species in the Field With Deep Learning and Drone Technology
Short description of the practice	Commercial drones integrated with deep learning models detect invasive plants in the field, improving precision and enabling real-time monitoring.
Keywords	aerial imagery, deep learning, drones, invasive plants, U-Net, remote sensing
Organisation in charge of the good practice	Department of Computer Science, Rhodes University, Grahamstown, South Africa
Implementation level of the practice	Level: Regional/National Country: South Africa Region: Eastern Cape (Grahamstown area) City: Grahamstown
Website	https://doi.org/10.1111/2041-210X.13473?utm_source=chatgpt.com
Detailed information on the practice	Traditional plant detection studies rely on orthomosaics and off-board processing, creating delays and limiting real-world use. This practice introduces a deep learning U-Net segmentation model integrated with a DJI Mavic Pro drone to detect invasive alien plants (Hakea genus) directly in the field. Datasets were collected from Eastern Cape, South Africa, and annotated for training. Model training used data augmentation (flips, rotations, brightness variation) and weight map-based loss to handle uncertainty in plant edges. Integration with a custom Android app allowed real-time segmentation overlays during drone flights. Stakeholders: Rhodes University, conservationists, land managers in the fynbos biome. Beneficiaries: ecological researchers, biodiversity managers, and invasive species control programs.
Timeframe	Dataset collected over two months; field trials conducted in early winter. The methodology is adaptable for ongoing monitoring.
Approximate cost	Not specified. Involves DJI Mavic Pro drone, Android devices, annotation labor, and computing resources for model training. Lower cost than satellite/hyperspectral approaches.
Results achieved	Final model achieved F1-score of 83% and accuracy of 96% on the test set. Field trials achieved recall of 75% with in-flight detection capability. Brightness augmentation improved F1-score by 27%; weight map-based loss improved precision by 15%.
Potential for learning or transfer	Demonstrates that real-time deep learning detection on drones is feasible for invasive species monitoring. Transferable to other species or environments where rapid, on-site detection is critical. Offers potential for intelligent drone interactions (closer inspection, bio-control delivery). Limitations: high false positives, seasonal variation effects, and dependence on annotated training data.
Additional material	Full article: James, K., & Bradshaw, K. (2020). Detecting Plant Species in the Field With Deep Learning and Drone Technology. Methods in Ecology and Evolution 11:1509–1519.
Contact person	Name: Katherine James Affiliation: Department of Computer Science, Rhodes University Email: katherine.mf.james@gmail.com