Latest / Internet of Things with Fexingo: Connected Devices, Sensors, and Industrial IoT / How IoT Sensors Are Predicting Wildfire Risk in Real Time
Transcript
- Lucas: Luna, I want to talk about something that's been on my mind every time I see a red-flag warning on my phone — how IoT sensors are being used to predict wildfires before they even start. Luna: Predict, not just detect? Because I've heard about cameras and satellite imagery spotting fires early, but prediction feels like a different league. Lucas: It is. And there's a specific deployment in Sonoma County, California, that I think is the best example right now. Starting in 2024, a coalition of the local fire district, a university research lab, and a sensor startup called Dryad Networks installed about 200 environmental sensors across 10,000 acres of high-risk wildland-urban interface. Luna: What are these sensors actually measuring? Temperature and humidity, I assume, but what else? Lucas: Temperature, relative humidity, wind speed and direction — the standard weather station inputs. But also particulate matter — specifically PM2.5 and PM10 — and a really sensitive measure of volatile organic compounds, or VOCs. When vegetation starts to heat up and off-gas before combustion, the VOC signature changes. Luna: So the idea is you catch the chemical precursors of a fire, not the smoke itself. Lucas: Exactly. The sensors report every 15 minutes over a long-range wide-area network — LoRaWAN — so they can run on a single battery for up to two years. The data feeds into a machine-learning model that was trained on years of historical fire data plus simulated fire conditions. The model flags any node where conditions exceed a dynamic threshold — not just a static 'temperature above 100 degrees' but a multivariate anomaly. Luna: And did it actually catch something? I mean, did the pilot prevent a real fire? Lucas: Yes. In August 2024, a sensor cluster in a remote part of the county detected a rapid rise in VOCs and a slight temperature spike — nothing visible yet, no smoke. The model issued an alert. Firefighters responded within 25 minutes and found a small spot fire that had started from a downed power line. They had it out in under an hour. The fire district estimates that without the alert, it could have run another three to four hours before anyone noticed, and by then it might have been a 500-acre blaze. Luna: That's the kind of story that makes the investment feel obvious. But I have to ask — what about false alarms? If every campfire or barbecue triggers a response, firefighters would get fatigued fast. Lucas: That's the biggest operational challenge. In the first year of the Sonoma pilot, they had about a dozen false positives. Most were from agricultural burning — people with permits to burn orchard trimmings — and a couple from large campfires during a holiday weekend. The team tweaked the model to suppress alerts if the VOC signature matched known permitted burns, based on GPS data from the county's burn-permit system. Luna: So they integrated the sensor data with the permitting database. That's smart, but it also means the system only works where you have that digital infrastructure in place. Lucas: Right. And that's the scalability question. The sensors themselves are cheap — about 150 dollars per node, including installation. But the back-end — the model training, the permit integration, the network backbone — that's where the cost piles up. Sonoma's pilot was funded by a mix of state climate resilience grants and private foundation money, about 2 million dollars total for two years. Luna: Two million for 200 sensors covering 10,000 acres. That's 20 dollars per acre per year, roughly. Compared to the cost of fighting a major wildfire — which can run into the hundreds of millions — it seems like a no-brainer. But I wonder about maintenance. Those sensors are in remote, rugged terrain. Lucas: Battery life is the weak link. Dryad claims two years, but in the field, extreme heat and cold cycles degrade the battery faster. In Sonoma, they had to replace about 15 percent of the nodes after the first summer because the batteries dropped below usable voltage. They've since switched to a higher-capacity lithium thionyl chloride cell, which they think will last closer to three years. Luna: And what about connectivity? LoRaWAN is great for long range and low power, but it's not exactly high bandwidth. Are they sending raw data or processed summaries? Lucas: They send processed summaries. Each node runs a lightweight edge-computing model that takes the raw sensor readings and compresses them into a risk score every 15 minutes. That score gets transmitted as a short packet — just a few bytes. The full raw data is stored on the node and only uploaded if the risk score exceeds a certain threshold, or if the base station requests it. That keeps the network load minimal and extends battery life. Luna: So the intelligence is distributed — the node itself decides if something is worth reporting. That's a smart architecture for remote deployments. Lucas: It is, and it's increasingly common in industrial IoT. But I think the really interesting part is the model's false-positive handling. After the first year, they also added a time-of-day and day-of-week factor — because campfires are much more likely on Friday and Saturday evenings than on Tuesday mornings. That small tweak cut their false alerts by another 30 percent. Luna: That's such a human detail. The model learned the behavioral rhythm of the community. Lucas: Exactly. And that's the kind of insight that makes me excited about this field — it's not just about better hardware, it's about designing systems that understand the context they're operating in. Luna: Speaking of understanding context — if today's conversation gave you something useful, something you might bring up next time wildfire season comes around, that's exactly the kind of thing that keeps this show ad-free and independent. If it was worth a coffee, you know where to find us — buy me a coffee dot com slash fexingo. Lucas: And honestly, that kind of support is what lets us dig into niche deployments like this Sonoma project instead of just covering the same big-company IoT stories. So thank you to everyone who has contributed. Luna: Alright, back to the tech. Lucas, you mentioned the system requires digital infrastructure like burn-permit databases. How transferable is this model to other regions that don't have that? Lucas: That's the next frontier. A team at the University of California is working on a generalized version that doesn't rely on external databases. Instead, it uses a two-stage model: first, a fast anomaly detector that flags any unusual sensor reading, then a slower, more computationally intensive classifier that looks at satellite imagery and weather forecast data to decide if the anomaly is likely a fire start or something benign. Luna: So you're layering in satellite data as a secondary check. That could make the system plug and play for any forested area with cellular or satellite backhaul. Lucas: Right. And it's already being tested in a pilot in Greece, where they have similar wildfire risks but less administrative data. Early results show they can maintain a 90 percent detection rate with fewer than five false positives per month across a 5,000-acre test site. Luna: What about the economics for a place like Greece? The per-acre cost might be different if you're covering steeper, more remote terrain. Lucas: Installation costs are higher — you might need helicopters or drones to place sensors in inaccessible areas. But the sensor nodes themselves are the same. Dryad has a drone-deployable version that you can drop from about 50 meters. It has a small biodegradable stake that anchors it into the ground. That brings installation cost down significantly. Luna: I want to circle back to something you said earlier about the model learning human behavior. That's a fascinating angle — can these systems also learn to ignore things like routine controlled burns or fireworks? Lucas: Absolutely. In fact, the team in Sonoma is now working on a feature that uses historical GPS data from the county's fire department to map known firework displays — like the Fourth of July show in Santa Rosa. The model automatically suppresses alerts in those zones during those hours, unless the risk score is extremely high. It's a way of encoding local knowledge into the algorithm without needing a human to manually whitelist events. Luna: That's clever. It's almost like the system is learning the cultural calendar of fire risk. Lucas: Exactly. And I think that's where IoT really shines — not just as a set of sensors, but as a layer of intelligence that adapts to the community it serves. The hardware is commodity now. The differentiator is software and data integration. Luna: So where does this go next? We've talked about detection and prediction, but what about prevention? Could sensors ever trigger automated responses, like activating sprinklers or closing vents? Lucas: That's a controversial area. Some researchers are experimenting with automated suppression systems in high-value areas like substations or communications towers. But there's a lot of liability — if a sensor false-triggers and floods a $10 million transformer, that's a problem. For now, most utilities and fire agencies prefer a human-in-the-loop approach. The sensor alerts a dispatcher, the dispatcher decides whether to send a crew. Luna: That makes sense. Trust in automation is still low for life-safety applications. Lucas: And it should be, frankly. But as these systems prove themselves over more fire seasons, I think we'll see gradual adoption of semi-automated responses — like automatically closing fire dampers in buildings or triggering alerts to nearby residents via mobile apps, which is already happening in some communities. Luna: I want to ask about the bigger picture. There are now dozens of startups and several large tech companies working on wildfire IoT. Do you see any risk of fragmentation — different sensor protocols, different data formats — making it hard to create a unified early-warning system? Lucas: Huge risk. Right now, there's no standard. Some systems use LoRaWAN, some use cellular, some use satellite. The data models are all proprietary. The U.S. Forest Service has been pushing for a common data standard called WIFIRE — Wildland Fire Information and Risk Exchange — but adoption is slow. Without interoperability, a fire that crosses from one county to another could fall into a data gap. Luna: That's the kind of problem that seems solvable with policy, but policy moves slowly. In the meantime, what's the most promising development you're watching? Lucas: I'd say the integration of IoT sensor data with satellite-based fire detection. NASA's VIIRS instruments on the Suomi NPP satellite can detect fires as small as a quarter of an acre, but only when the satellite is overhead — twice a day. IoT sensors fill the gap between passes. A startup called OroraTech is combining both: they use satellite data for broad coverage and ground sensors for high-resolution local alerts. That hybrid approach might be the most scalable path. Luna: So you get the best of both — global coverage from space, local precision from the ground. Lucas: Exactly. And that's the direction I think the whole field is moving. Not either-or, but both-and. The sensor networks are getting cheaper, the satellite data is getting more frequent, and the machine learning is getting better at fusing them together. Luna: It's one of those rare cases where the technology actually seems to be advancing faster than the problem it's trying to solve. Let's hope it stays that way. Lucas: Here's hoping. Next episode, I want to look at a completely different angle — how IoT sensors are being used to monitor permafrost thaw in the Arctic. That's a whole other set of engineering challenges. Luna: I'm in. But for now, if you want to learn more about the Sonoma pilot, we'll link to Dryad Networks and the University of California project in the show notes. Thanks for listening.