WeedElec: Combining Artificial Intelligence, Drones, and Robotics for Herbicide-Free Weeding
The WeedElec project explores a chemical-free precision weeding solution. Its principle: locate undesirable plants using images captured by drones and a robot, identify them using artificial intelligence, and eliminate them individually via a controlled electrical discharge. Pl@ntNet technologies and data play a crucial role in this system.
In an agricultural field, plants that compete with crops are referred to as weeds. Their management still relies heavily on the use of herbicides, which impacts water quality, ecosystems, and human health. Reducing these treatments requires developing methods capable of automatically distinguishing crops from weeds, and intervening only where necessary.
Supported by the French National Research Agency (Agence nationale de la recherche – ANR) as part of the ROSE Challenge 2017, the WeedElec project brings together specialists in agricultural robotics, imaging, artificial intelligence, botany, and plant physiology. Its objective is to build the scientific and technical foundation for an integrated system combining a drone, an agricultural robot equipped with a high-speed articulated arm, and a high-voltage electrical weeding tool.
In this project, the combined analysis of sensor images makes it possible to map infested areas, anticipate competition risks between crops and weeds, and plan the path of autonomous weeding tools. These tools can then avoid unnecessary movements, thereby reducing intervention time, energy consumption, and soil compaction.
Identifying Weeds Using Pl@ntNet
Recognizing a plant photographed by a person and automatically detecting a small weed within a crop are two very different challenges. Images captured by a drone or an onboard camera can feature very small plants that are partially hidden or viewed from unusual angles. They are also subject to variations in lighting, weather, soil conditions, and growth stages.
The teams involved in Pl@ntNet investigated adapting plant recognition methods to these agricultural conditions. The project focused in particular on deep learning—a family of artificial intelligence techniques capable of learning to recognize visual characteristics of species from vast collections of images.
Several approaches were explored:
- Adapting these models to images automatically captured by drones and robots;
- Developing a workflow to run them on a robot with limited computing power.
This research helps expand the applications of Pl@ntNet. The platform is no longer used solely to identify photographs submitted by its users; its methods can also be adapted to automated sensors and integrated into equipment capable of taking direct action in the field.
Intervening Exclusively on Target Plants
Once the weeds are located, the robot navigates to the relevant zones. Its onboard camera refines the detection in real time to accurately position the weeding tool. The tool then applies a high-voltage electrical pulse to the targeted plant.
The passage of the electric current damages plant tissues and can reach the root system. However, not all species react the same way. The effectiveness of the treatment depends notably on the size and growth stage of the plant, its water status, its root architecture, as well as the moisture level and type of soil.
WeedElec therefore studied the electrical signature of various weeds, crops, and soil categories. The goal was to determine the signal characteristics best suited to each situation: voltage, duration, frequency, and energy amount. This work aimed to achieve a lethal effect with the lowest possible energy consumption, while preserving the cultivated crops.
Research in Service of Agroecology
By combining aerial detection, onboard recognition, and localized electrical intervention, WeedElec offers an alternative way to manage weeds: observe before acting, identify the plant, and intervene only on the target.
The complete solution was designed to be tested in field crops and market gardening plots, particularly on maize (corn) and carrots. Its various components can also be leveraged independently: identification algorithms, hyperspectral analysis methods, the electrical tool, or the robotic system.
For Pl@ntNet, WeedElec has served as a key research ground at the intersection of botany, artificial intelligence, and precision agriculture. The project has helped make automated plant recognition more robust when faced with new observation conditions and demonstrates how knowledge generated by a participatory platform can support the development of agricultural tools that reduce the need for herbicides.
- Program: ROSE Challenge, 2017 Edition
- Funding: French National Research Agency (Agence nationale de la recherche)
- Fields: Agroecology, Artificial Intelligence, Plant Recognition, Hyperspectral and Color Imaging, Drones, Agricultural Robotics, Electrical Weeding
- Website: https://www.challenge-rose.fr/projet/weedelec/
