TLDR: A new robotic system uses AI (YOLO11n for detection, YOLO11n-seg for segmentation) on an NVIDIA Jetson Orin Nano to detect weeds and their canopy size in real-time. It then precisely adjusts herbicide spray volume using an Arduino-controlled nozzle system. Indoor trials with hibiscus plants showed it effectively varied spray coverage based on plant size, achieving an average of 24.22% coverage and demonstrating a clear increase in spray volume for larger canopies, thereby reducing chemical waste and environmental impact.
Modern agriculture faces significant challenges with traditional herbicide application methods, which often lead to high costs, environmental pollution, and the development of herbicide-resistant weeds. To combat these issues, researchers have developed an innovative robotic system that uses artificial intelligence (AI) for real-time weed detection, intelligent spraying based on plant canopy size, and evaluation of spray patterns.
This new system is designed to make herbicide application more precise and efficient. It integrates a vision-guided, AI-driven variable rate sprayer that can identify weeds, estimate their canopy size, and adjust nozzle activation in real time. The core of the system relies on lightweight deep learning models, YOLO11n for detecting weeds and YOLO11n-seg for segmenting their canopy, deployed on an NVIDIA Jetson Orin Nano for on-board processing. An Arduino Uno-based relay interface then controls solenoid-actuated nozzles, ensuring that only the detected weed areas receive herbicide, and the amount sprayed is proportional to the weed’s size.
System Components and Operation
The robotic sprayer is mounted on a Farm-ng Amiga platform, a robust electric robot designed for agricultural research. This platform carries the necessary tanks, control units, and embedded processors. A custom aluminum spray boom, equipped with four nozzles spaced 30 cm apart, is attached to the robot. Each nozzle has a 12VDC PWM actuated solenoid valve, allowing for individual and precise control of spray output. A 50-liter tank supplies the liquid, pressurized by a diaphragm pump.
The AI-driven embedded control unit, housed in a weather-resistant enclosure, manages all perception, decision-making, and actuation tasks. The NVIDIA Jetson Orin Nano processes live image data from front-facing RGB cameras, running the AI models to detect weeds and calculate their canopy area. An Arduino Mega microcontroller acts as an intermediary, receiving commands from the Jetson and translating them into activation signals for the spray nozzles. This setup allows for fine-grained spray modulation, adjusting spray durations based on the estimated weed canopy area.
AI Models and Performance
The YOLO11n model, used for weed detection, achieved impressive results in indoor trials, with a mean Average Precision (mAP@50) of 0.98, precision of 0.99, and recall near 1.0. This indicates high accuracy in identifying weeds. The segmentation model, YOLO11n-seg, which is crucial for estimating canopy size, achieved a mAP@50 of 0.48, precision of 0.55, and recall of 0.52. While the segmentation performance is lower than detection, it is still effective for guiding variable spray actuation.
Canopy-Aware Spraying Validation
Indoor trials were conducted using 15 potted Hibiscus rosa sinensis plants of varying canopy sizes (small, medium, and large) to simulate diverse weed patches. The system’s performance was validated using water-sensitive papers, which change color upon contact with water, allowing for precise quantification of spray coverage. The results showed an average spray coverage of 24.22% in canopy-present zones. Crucially, the system demonstrated its ability to dynamically modulate spray intensity: mean spray coverage increased from 16.22% in small canopies to 21.46% in medium and 21.65% in large canopies. This clear upward trend confirms the system’s effectiveness in adjusting spray volume based on the detected canopy size, minimizing overspray on smaller weeds and ensuring adequate coverage for larger ones.
Also Read:
- AI in Agriculture: Empowering Farmers Through Knowledge Beyond Yield Optimization
- Optimizing Autonomous Driving AI: How Smart Action Choices Lead to Safer, Faster Learning
Future Directions
This research highlights the feasibility of integrating real-time deep learning with low-cost embedded hardware for selective herbicide application. Future work will expand the system’s detection capabilities to include three common South Dakota weed species: waterhemp, kochia, and foxtail. The system will also undergo further validation through both indoor and field trials in soybean and corn cropping systems. The ultimate goal is to refine segmentation accuracy, reduce nozzle actuation latency, and streamline image acquisition for faster processing, paving the way for smart, sustainable spraying technologies that reduce chemical usage and enhance targeting precision in agriculture. For more detailed information, you can refer to the full research paper available at arXiv.org.
