Enhancing a visual crop detection model for an agricultural firm

Developing an advanced ML model optimized for edge processing

OneSix collaborated closely with the agricultural firm’s internal teams to understand the unique challenges and nuances of autonomous weeding, particularly the need to balance high-speed processing with accuracy on edge devices.

To address these requirements, we designed an optimized machine learning model that leverages state-of-the-art techniques in object detection, enabling it to run quickly on edge hardware without sacrificing accuracy. Through careful tuning, we achieved a model that could process images faster than real-time and maintain, or even exceed, the detection accuracy of the original model.

In addition to building the core model, OneSix broadened the project’s scope to introduce additional features that further increased the robustness and versatility of the robotic weeding solution:

• Multi-Crop Detection: The team transitioned from a single-crop detection model to a multi-crop detection framework, allowing the same model to be deployed across fields with different crops. This innovation simplifies deployment for end users by enabling robots to operate in diverse environments without needing separate configurations for each crop type. As a result, operators can now deploy the robots across fields, regardless of the crop variety, significantly streamlining operations and reducing setup times.
• Unannotated Image Search: To optimize training, OneSix implemented an unannotated image search feature, allowing the model to automatically explore large datasets of untagged images. This feature enabled the identification of complex crop and weed combinations, which helps improve detection accuracy by training on a wider variety of real-world data, all without requiring extensive manual annotation efforts.
• Generative Image Training: Adapting machine learning models to the variable conditions of agricultural environments—such as differences in lighting, soil texture, and camera angles—was essential for real-world deployment. To meet this challenge, OneSix introduced generative image training using advanced AI-driven image translation techniques. This approach simulates a variety of field conditions in a controlled setting, producing synthetic images that allow the model to learn from diverse scenarios without needing new data collection. By leveraging generative AI, the model could continuously improve its adaptability, providing reliable performance in various field conditions.

Results

Increased speed, accuracy, and adaptability

The optimized crop detection model delivered substantial improvements in both speed and accuracy, meeting the demands of real-time, edge-based processing. The integration of multi-crop detection, unannotated image search, and generative image training provided significant benefits, allowing the autonomous robots to perform reliably across different crops and environments. By automating complex weeding tasks, the model enables the agricultural firm to reduce manual labor costs and minimize pesticide use, supporting sustainable farming practices.

With this advanced, adaptable machine learning platform, the agricultural firm is now equipped to deploy its robotic weeding solution at scale, enhancing operational efficiency and promoting environmentally friendly practices in the agricultural sector. The success of this project highlights the transformative potential of machine learning in addressing complex agricultural challenges, showcasing how AI-driven innovation can drive meaningful impact in real-world applications.