By Blake Jackson
Researchers at the University of Missouri have created FieldVision, an artificial intelligence framework designed to help groups of agricultural drones determine where image-processing tasks should be completed.
The system enables drones to decide whether to analyze data onboard, send it to a nearby edge computing server, or transfer it to cloud-based resources.
Simulation experiments showed that FieldVision performed better overall than conventional rule-based strategies and approaches designed for individual drones.
The framework increased rewards reduced missed processing deadlines and improved system reliability. It also allowed drones to make decisions independently without requiring direct communication with other drones during missions.
The research team included Mizzou investigators Andrew Hellman, Bishwas Wagle, Alicia Esquivel Morel, Juan Mogollon, Jianfeng Zhou, Kannappan Palaniappan and Prasad Calyam.
Sean Peppers of Florida Gulf Coast University, Vincent Zheng of Stony Brook University and Arunava Roy of the University of Memphis also contributed.
The project received support from the National Science Foundation through a Research Experiences for Undergraduates grant focused on consumer networking technologies.
Agricultural drones can capture large volumes of aerial imagery for applications such as crop counting, crop-health monitoring, anomaly detection, and field inspections.
However, turning that imagery into useful information quickly can be difficult because drones have limited onboard computing and battery resources. Rural wireless networks can also fluctuate, making data transfers to edge servers or cloud platforms less predictable.
These challenges increase when several drones operate simultaneously. Multiple aircraft may compete for wireless bandwidth and shared computing capacity, meaning a decision that benefits one drone could increase congestion and processing delays for others.
FieldVision uses multi-agent reinforcement learning to address these competing demands. Through centralized training with decentralized execution, drones learn during training how shared network and computing resources affect performance.
Once deployed, each drone can independently select a processing location using information available to it locally.
“Our goal is not simply to make an individual drone more productive. We want multiple drones to learn how to use limited network and computing resources effectively as conditions change, without depending on continuous communication among the drones.” said Bishwas Wagle.
The technology could benefit farmers, agricultural researchers, and organizations using drones for precision agriculture by helping deliver processed imagery faster during field operations.
“Although we developed FieldVision around precision agriculture, the underlying problem is much broader. It solves for how a group of autonomous systems can intelligently share computing and networking resources when the environment is constantly changing.” said Bishwas Wagle.
Photo Credit: gettyimages-seregalsv
Categories: Missouri, Education