Wildfire Sentinel

Wildfire Sentinel is a solar-powered camera and thermal node that detects wildfire smoke and ignitions entirely on an Arduino UNO Q. It runs on SuperFloat, an exponent-free number format that makes CNN inference integer-only. It sounds a local alarm with no network at all, queues alerts until any link comes back, and never sends an image anywhere.

Track A (Developers). Challenge areas: 1 Anticipate and Mitigate Risk, 3 Enable Real-Time Response, 5 Operate in Mission-Critical Environments.

Code, hardware files and tests: https://github.com/aloshdenny/superfloat.arduino (MIT)

Wildfires are cheapest to stop in their first minutes and most destructive once they run. The Camp Fire (2018), Lahaina (2023) and the Los Angeles fires of January 2025 are recent reminders: in each, time to detection and time to warning shaped the losses.

The places where ignitions start are where detection is weakest:

• wildland–urban interface towns, ranches, tribal lands, utility corridors and remote infrastructure;
• often beyond reliable cellular coverage;
• often reliant on someone noticing smoke and calling it in.

Camera networks exist, but the mature ones stream video to cloud AI. That needs:

• continuous high-bandwidth backhaul;
• per-camera subscriptions that small counties, fire districts and tribes struggle to fund;
• links that stay up. A fire damages power and communications first, so the systems are weakest when they matter most.

The affected community is rural and wildland–urban interface residents in fire-prone regions of the US, and the small, often volunteer, fire services that protect them.

• Volunteer and county fire departments, and tribal fire programs. They get minutes-scale alerts with a bearing (the tile index maps to a direction) for the price of a node, not a subscription.
• Residents of WUI communities, ranches and remote homes. A local buzzer and LED alarm work even when the network does not, and webhook alerts can feed existing community alerting.
• Utilities and operators of remote infrastructure (substations, pumping stations, telecom sites). They can watch their own assets and corridors without streaming video off site.
• Land managers and researchers. The models, firmware and data pipeline are open source and adaptable to other hazards, such as smoke from industrial sites or post-fire hotspot monitoring.

Hardware.

• Arduino UNO Q: Qualcomm Dragonwing QRB2210 with 4x Cortex-A53 on Debian, plus an STM32U585 Cortex-M33.
• A USB camera.
• An MLX90640 32x24 thermal array on Qwiic.
• A buzzer, a mute button and the on-board LED matrix.
• A 60 W solar panel with a 320 Wh LiFePO4 battery, about 3 days of autonomy.
• A 3D-printed mount in an IP65 box.

All hardware design files are in the repository.

Two tiers on one runtime.

• Camera tier (QRB2210). Each frame is resized and cut into 15 tiles of 128 px, following the tiled approach of SmokeyNet (Dewangan et al., 2022).
  • Each tile is classified clear, smoke or flame by SmokeNet: 28 M MACs and 269 KiB of weights.
  • Four threads run on the four A53 cores.
  • A temporal voter requires 4 of 6 frames for smoke and 2 of 3 for flame, which suppresses glare, cloud and dust.
• Thermal tier (STM32U585). EmberNet (0.66 M MACs, 15 KiB) classifies MLX90640 frames together with their frame-to-frame difference, so flicker counts as evidence.
  • It runs on the MCU, independently of Linux.
  • It keeps watching day and night, even if Linux hangs; a heartbeat watchdog shows the fault on the matrix.

The edge AI method: SuperFloat.

SuperFloat (my prior work, aloshdenny/superfloat) drops the exponent entirely: a value is a sign plus fraction bits, i.e. Q1.(x−1) fixed point. Trained network weights almost never leave [−1, 1], and my benchmarks show SF8 weights matching FP32 on EuroSAT remote-sensing classification.

Because every SuperFloat scale is a power of two, a whole CNN runs on integers:

• int8 multiplies and int32 accumulation;
• one rounding shift and a clamp per layer;
• no zero points, no per-channel float scales, no requantisation multipliers.

SF4 weights (2 per byte) are a subset of SF8 and use the same kernel.

The quantisation-aware training is constructed so that the PyTorch forward pass is bit-for-bit identical to the C runtime on the board. This is checked by the test suite, so the accuracy measured in training is the accuracy deployed.

Edge Impulse.

• Dataset management: tiles from D-Fire and HPWREN, then our own field captures, are uploaded and labelled in Edge Impulse.
• Baseline comparison: each trained model is exported as an ONNX float twin for Edge Impulse BYOM. Edge Impulse profiles it and builds an int8 .eim for the UNO Q (runner-linux-aarch64).
  • SuperFloat and Edge Impulse int8 are then compared on the same board and the same tiles, for accuracy, latency and power.
• Integration: the app runs in Arduino App Lab, using the Bridge, Camera and WebUI bricks.

• Connectivity fails first.
  • Many ignition-prone areas have little or no coverage.
  • Fires damage the power and communications infrastructure cloud systems depend on.
  • An edge node decides locally. It sounds a local alarm with no network, and queues alerts to disk until any link returns.
• Bandwidth and cost.
  • Streaming 720p around the clock is gigabytes per camera per day, which is prohibitive over cellular or satellite and on solar power.
  • An alert is about 200 bytes. Moving the decision to the board cuts backhaul by roughly six orders of magnitude.
• Latency. The system catches smoke within seconds of a frame, without queueing behind a cloud pipeline.
• Privacy.
  • Frames never leave the board and are never stored.
  • The model has no person class.
  • Alerts carry a tile index and a score, not an image.
  • This keeps the project clear of surveillance concerns and makes community consent much easier.
• Power. Integer-only inference on the A53 and the M33 fits a solar budget. The MCU tier keeps watching even when the Linux side is throttled.

• Open source from day one: the code, the trained models, the hardware files and the test suite are public under MIT, with commit history showing how it was built.
• devEco project page: weekly updates through stage 2, with on-board measurements (latency, power, accuracy) as soon as the kit arrives, including the ones that disappoint.
• Edge Impulse public project: the labelled field dataset and the int8 baseline, so others can reproduce the comparison.
• Write-ups:
  • an adaptation guide (other hazards, other cameras, other boards);
  • a technical article on integer-only SuperFloat inference;
  • a demo video of the node detecting smoke on recorded HPWREN sequences and a controlled burn-barrel test.
• Negative results get published too. The SuperFloat benchmark repository already documents where the format fails, such as transformer activations. This project will do the same.

Already working (before the kit arrives):

• The SuperFloat integer runtime: C99, no allocator, NEON on aarch64.
  • It is bit-exact against a reference implementation and against the training forward pass.
  • It is tested on macOS arm64 and Linux aarch64, with about 120 automated tests.
• The QAT pipeline and exporter, the SmokeNet and EmberNet definitions, the full App Lab application, and the MCU sketch.
  • The sketch is compiled and exercised on a host against stub Arduino APIs.
  • Those host tests cover the watchdog, alarms, mute and a thermal hotspot detected by EmberNet.
• Edge Impulse tooling, the BOM, wiring, the power budget and the printable mount.

Not yet proven, and the plan for each:

• A53 speed.
  • Projected at 15–30 ms per tile and about 0.1 s per 15-tile frame on four cores. A frame every 2 s leaves a 10–20x margin.
  • Fallbacks: a narrower SmokeNet or fewer tiles.
  • First measurement in week one with the kit.
• Accuracy on real smoke.
  • Neither model has been trained on real imagery yet. The thermal model is a bootstrap trained on synthetic frames.
  • SmokeNet training on D-Fire is a stage 2 week-one task.
  • Success means meeting or beating the Edge Impulse int8 baseline on the same tiles.
• False alarms. These are the reason fire agencies distrust automated detection. The temporal voter will be tuned on real HPWREN sequences, and false alarms per camera-day will be reported alongside time to detect.
• Library and toolchain unknowns on the UNO Q Zephyr core, such as the MLX90640 driver. Fallback: port the Melexis reference C driver.
• Night. RGB smoke detection does not work in darkness. The thermal tier covers ignitions at night, and that limitation is stated plainly.

Scope and support. I am a solo developer, and the scope is sized for one person over seven weeks: everything except on-board measurement and field data is built. The main risks are performance and data, not missing software.

The deliverable for December 2 is:

• one working outdoor node;
• measured accuracy, latency and power against the Edge Impulse baseline;
• a documented path to more nodes.

• Repository: https://github.com/aloshdenny/superfloat.arduino
• SuperFloat format and benchmarks: https://github.com/aloshdenny/superfloat
• Atreides, a SuperFloat (SF16) accelerator in Sky130: https://github.com/aloshdenny/superfloat.gpu