Smoke, fog or dust? Why fire detection systems raise false alarms

Why the lookalikes are difficult

A normal video camera records appearance. It does not directly measure what a cloud is made of. Smoke, drifting fog and airborne dust can all create moving, partly transparent patches that obscure the background.

A study of wildfire camera networks identifies clouds, haze, fog and dust as difficult distractions when early smoke is thin and low in contrast. Colour alone is not enough. A pale patch might be harmless, but pale smoke can also be a sign of fire.

Some misleading patterns come from the equipment’s immediate surroundings. In a field study in Portugal, researchers documented false detections caused by water drops on the lens, sun reflections and even a pole. Improving reliability therefore involves the camera installation as well as the AI.

The cost of a quieter system

Detection software often gives a suspected plume a score. An alert is triggered when the score passes a chosen threshold, sometimes with additional checks. Our guide to how AI detects wildfire smoke explains that sequence.

For the same model, raising the threshold generally reduces false alerts. But it can also reject real smoke with a weaker score. Lowering it has the opposite trade-off. A quiet system is not automatically a good detector.

Repeated false alarms have a human cost too. An account of operating a public wildfire detection system describes how investigating them can reduce the perceived urgency of the next notification. The aim is to preserve attention without filtering away useful warnings.

Why accuracy can mislead

A headline accuracy percentage can hide this trade-off. Google’s guide to evaluating classifiers explains that accuracy can be misleading when one kind of example greatly outnumbers the other.

Imagine a test containing 1,000 images, just one of which shows smoke. A system that always answers “no smoke” gets 999 right. Its accuracy is 99.9%, yet it misses the only smoke image. This is an invented example, not a result from a detection product.

Better questions separate the outcomes: how much real smoke is detected, how many alerts are false, and how long detection takes. These measures describe different parts of performance; none should be read alone.

Reducing false alarms in practice

Several checks can work together. Training can include verified examples of local lookalikes. A sequence of images can help a reviewer see how a suspicious patch develops. Camera maintenance can address recurring obstructions or water drops.

The HPWREN operational account describes using confirmed false detections to improve training and time-lapse images to help people distinguish smoke from distractions. This is an ongoing process: a successful demonstration does not establish performance in every season or location.

It also matters where an alert goes. A machine flag, an operator-confirmed incident and a request for an emergency response are different stages. In ALERTCalifornia’s workflow, trained watchstanders vet potential incidents before confirming them. Ask which stage a supplier’s false-alarm figures describe.

What should a site manager ask?

Ask for false alerts per camera per day, alongside missed detections and detection time under the same settings. Check whether repeated flags for one event count separately. Request examples from difficult weather, lighting and site activity similar to your own.

Agree who checks each alert and what happens next. Previous false alarms are not a reason to dismiss a new one without assessment. Our guide to using existing security cameras for fire detection explains how to start assessing the camera network itself.