Detecting and Mapping Stoebe vulgaris Post Herbicide Control, Using UAV Imagery and Machine Learning

Stoebe vulgaris is a declared indigenous bush encroacher species in South Africa. It has invaded over 11 million ha of grasslands. It is commonly called bankrupt bush due to its ability to outcompete other indigenous forb and grass species, decreasing grazing capacity, biodiversity, and ecosystem functioning, eventually leading to financial ruin for farmers. Landowners are legally required to control the plant. A herbicide trial was set up in a severely encroached camp at Dundee Research Station in KwaZulu-Natal, South Africa, to test the effectiveness of metsulfuron-methyl (50 g active ingredient ha−1) in controlling S. vulgaris. Applying metsulfuron-methyl provided a significant long-term reduction in S. vulgaris cover over six years. However, effective monitoring and management strategies depend on knowledge of the spatial distribution and expansion patterns of invasive species. We evaluated the ability of UAV-based imagery and machine learning, using Picterra, to detect and map S. vulgaris, while determining the optimal parameters to maximise detection accuracy. The best season for image acquisition was late summer when vegetation was at peak growth and maturity, providing the best spectral distinction between species, under light overcast and mild wind conditions. We recommend careful consideration of the flight orientation to the solar angle. We achieved 92% detector accuracy, with multispectral imagery enhancing the discrimination of similarly coloured plants. Plants smaller than 10 cm were not detected by the model. Our approach, using high-resolution drone imagery and AI, is capable of individual plant detection suited to a farm scale. This opens the way for using advances in drone technology for targeted, spot-application of herbicide.