In agriculture and horticulture, weeds are often removed with pesticides because they compete with the main plants for nutrients, water and light. However, not all weeds are harmful, they can even be beneficial. Researchers at the Technology and Support Centre (TFZ) in Straubing have been working on using artificial intelligence (AI) and drone images to apply herbicides in a more targeted and automated manner.

Broad database as a basis

The aim of the project was to develop an AI model that can distinguish crops from weeds in images. To achieve this, it was fed with a large amount of data: drones took over 121,000 images of sorghum fields in Straubing. The collected data was labelled by hand for the AI model which was ‘an incredibly time-consuming process’ according to project manager Michael Grieb. However, this is the only way the AI can learn what the heat-preferring energy crop sorghum looks like, which rows it grows in and how it can be distinguished from weeds.

Special training due to environmental influences

Wind, motion blur and direct sunlight were a particular challenge, as these had a negative impact on the image quality. This made it difficult for the AI to analyse the data, as the plants grow and change. The research group therefore had to additionally train the model for the different growth stages of the plant. With success: according to the TFZ, the AI can now differentiate between sorghum and weeds with almost no errors.

The foundations were laid in this research project together with the Weihenstephan-Triesdorf University of Applied Sciences, the Technical University of Munich, Straubing Campus for Biotechnology and Sustainability and the Bavarian State Research Centre for Agriculture. In a follow-up project, the technology is now to be integrated into field robots.

lh