EWIS2 Funded by Bavarian State Ministry for Food, Agriculture, Forests and Tourism

Project Description

Results

During this project, we generated drone captures of 81 distinct agricultural maize and sorghum fields in 12 bavarian couties. This dataset has a total of 300 ha and spans 2 years. We employed two different drones: The DJI Mavic 3M and a M300 with a P1 sensor. The drone images resulted in a Ground Sampling Distance (GSD) of 0.3 cm/px, which was precise enough to see small weeds between maize plants in BBCH 13.

We re-framed the weed detection task as “plant detection with maize removal”. Here, we detect every green pixel with standard computer-vision algorithms, i.e. HSV-based thresholding or Excessive-Green-Index. The AI part focuses on detecting the maize plants, which were removed from all green pixels in a consecutive step.

Before training an AI-based maize detector, the ground-truth data was labeled by hand using bounding boxes. This endeavour took months and more than 5 employees. In the end, it yield over 200'000 annotated maize plants, which were used to train different object detection models. 

The main bottleneck in AI-based weed detection is not the detection itself. It is the integration in the everyday life of farmers. Our models were able to detect maize in a variety of different agricultural settings and weather conditions. However, the process of geo-referencing drone images into an orthofoto took ages and was not feasible. Therefore, a method was adopted and further developed that allows individual drone images to be directly georeferenced and quickly processed into maps. 

The entire process - from drone flights to map creation and integration of the maps into a field robot, all the way through to site-specific mechanical weed control - was successfully implemented on a Robotti 150D (Agrointelli).