Latest / The Edge Computing Podcast with Fexingo: Local Compute, CDNs, and Distributed Infrastructure / How Edge Computing Is Unlocking Real-Time Agricultural Drone Spraying
Transcript
- Lucas: Luna, I've been looking at a number that keeps popping up in precision ag circles: a 47 percent reduction in herbicide use from a single field trial in Nebraska last season. And the enabler wasn't a better chemical or a new nozzle — it was edge computing on a drone. Luna: Forty-seven percent — that's not a tweak, that's a rewrite of the input budget. How does edge fit into a spraying drone? Lucas: The traditional approach is a drone flies a pre-programmed grid and sprays a uniform rate across the whole field. But weeds don't grow uniformly. You get patches of waterhemp, bare soil, and healthy corn all in the same pass. The cloud-dependent alternative — take a picture, send it to a server, get instructions back — is way too slow. By the time the drone has the answer it's flown past the weed. Luna: So the edge lets the drone make the decision in the air, in milliseconds. Lucas: Exactly. In the Nebraska trial they used a DJI Agras T40 retrofitted with an NVIDIA Jetson Orin NX module running a lightweight convolutional neural network. The onboard camera feeds frames at 30 frames per second, the model classifies each patch as weed, crop, or bare soil, and the spray nozzles adjust flow rate in real time. No round trip to the cloud. The whole inference loop is under 50 milliseconds. Luna: Fifty milliseconds — that's about the blink of an eye. But how accurate is that model? If it misclassifies a corn plant as a weed, you're spraying herbicide on your crop. Lucas: The model they deployed hit 94 percent accuracy on weed identification and 97 percent on crop classification. That's good enough for a pre-emergence pass where you're targeting early-stage weeds before they compete with the corn. The trade-off is you're running a smaller model than what a data center could handle — you compress the neural net to fit the Jetson's 20 watt power envelope. But they found that the speed advantage more than compensated for the slight accuracy drop versus a cloud model. Luna: So it's a classic edge trade-off: lower peak accuracy, but much better real-world performance because the latency disappears. What's the hardware cost for this setup? Lucas: The drone itself runs about fifteen thousand dollars. The Jetson module adds maybe fifteen hundred. So call it sixteen, seventeen thousand all-in. Compare that to a self-propelled ground sprayer — a John Deere or Case IH unit — which can run four hundred thousand dollars or more. For a farm that's a few thousand acres, the drone plus edge model starts to look very attractive on cost per acre, especially if you can run multiple drones in a swarm. Luna: But a ground sprayer covers more acres per hour, right? How does the throughput math work out? Lucas: It does. A single ground sprayer can cover about a thousand acres per day. A single agricultural drone does maybe three to four hundred. But you can deploy five drones for the price of one ground rig, and each drone is lighter — less soil compaction, less crop damage from driving through the field. The Nebraska farm used a team of four drones flying simultaneous passes, and their effective coverage rate was competitive with a single sprayer, plus they got the 47 percent herbicide savings. Luna: That's a compelling total cost picture. But there's another bottleneck I hear about: the FAA. Flying drones beyond visual line of sight — BVLOS — is still heavily restricted. How do you run a four-drone swarm if each one has to stay within sight of a pilot? Lucas: That's the regulatory hurdle. Right now most agricultural drone operations require a visual observer for each drone, which kills the economics of a swarm. But edge computing actually helps here too. The FAA is more open to BVLOS waivers when the drone has onboard detect and avoid capability — that is, it can sense obstacles and other aircraft locally without relying on a ground link. And that's exactly what the edge module enables. The same Jetson that does weed classification can run an object detection model for obstacle avoidance. Luna: So edge is solving both the ag problem and the regulatory problem on the same piece of hardware. That's neat. Do you have a sense of how many farms are actually doing this today versus just pilot trials? Lucas: It's still early — maybe a few hundred farms in the US, mostly in the Corn Belt and the Mississippi Delta. But the adoption curve is steep. A USDA survey from late 2025 showed that about 12 percent of row-crop farmers had used a drone for some kind of variable-rate application, and of those, roughly a third were using on-board edge processing. That's up from essentially zero three years ago. Luna: And the environmental angle — 47 percent less herbicide — that's the kind of number that gets regulators and sustainability officers interested. Speaking of keeping things sustainable without ads, I'm glad we can have these conversations without commercial interruptions. Lucas: Yeah, that's a deliberate choice on this show. We don't run ads because we think the content is cleaner that way — no sponsored segments masquerading as analysis. If you find episodes like this useful and want to support that ad-free approach, the link is buy me a coffee dot com slash fexingo. It's a small gesture that keeps the whole thing independent. Luna: Exactly. And it helps us keep digging into specific use cases like this one. So back to the ag economics — what's the next bottleneck? Battery life? Lucas: Battery life is a factor. A typical ag drone flies about 20 to 30 minutes per battery, so you need multiple batteries and a charging station in the field. But the edge module actually draws less power than a cellular radio transmitting full video to the cloud would. So edge processing extends flight time slightly. The bigger issue is data management — each drone generates about a terabyte of video per season. Where does that go? Luna: Does it all get processed on the edge or do you upload summaries later? Lucas: The current best practice is to store only the inference metadata — coordinates where herbicide was applied, classification logs — which is a few megabytes per flight. The raw video can be discarded. That's another edge advantage: you don't need to stream terabytes to the cloud. You just upload a tiny JSON file after each mission. Luna: So edge computing reduces bandwidth requirements too. I'm starting to think this is less a niche application and more a template for how precision ag will work across the board. Lucas: I think so. And the next step is combining multiple sensor modalities — adding multispectral or thermal cameras to the edge inference pipeline. A few startups are already testing drones that can distinguish between weed species and apply different chemistries per nozzle. That would push herbicide reduction even further, maybe to 60 or 70 percent. Luna: That's the kind of leap that could change the regulatory conversation around chemical inputs entirely. If you can prove you're only spraying what's needed, where it's needed, the argument for blanket restrictions weakens. Lucas: Right. And it all comes back to that fifty-millisecond inference loop running on a $1,500 edge module strapped to a drone. It's a small piece of hardware doing a lot of heavy lifting. Luna: Alright, next time I want to talk about the flip side: what happens when the edge model fails? How do you handle false negatives in the field? Lucas: That's a great angle. We'll dig into edge model robustness and fallback strategies. For now, thanks for listening to The Edge Computing Podcast with Fexingo.