Wildfire IoT Watch

Smoke plume seen from an HPWREN tower camera

Internet of Things & Big Data

Early Wildfire

Detection

Edge AI on an IoT camera network

Kanatbay Alibek · Shaykhmetov Alimzhan

KBTU · 2026

Smoke column rising behind a communications tower

The problem

Minutes matter

Satellite products such as NASA FIRMS report hotspots hours after a fire starts. A camera on a tower sees the smoke column almost at once.

Hours

satellite hotspot reports, depending on overpass and processing

6min

median time from ignition to alarm on our eight test fires

Goal

Recognise smoke in tower camera video, estimate where it is, and alert an operator with as few false alarms as possible.

Wildfire IoT02 / 10

IoT system

Architecture

LIVErunning in today's demo NEXTplanned

Camera

Fixed PTZ camera on a fire tower

Edge node

Raspberry Pi 5 or Jetson Orin Nano: YOLO smoke detector, 3-frame filter, bearing

LIVE · on laptop

MQTT · LTE

Server

MQTT broker, Python backend, triangulation

LIVE

TimescaleDB + PostGIS, MinIO, Grafana

Operator

Live map console and Telegram alerts

LIVE

Weather

Temperature, humidity, wind on the same tower

LIVE · simulated

NB-IoT

Wildfire IoT03 / 10

Edge first

Events, not video

233×

less uplink traffic: towers analyse frames on site and send only small events.

Video stream, 2 towers × 81 min48.4 MB

MQTT events, same period212.7 KB

MQTT topics

wildfire/

<tower>/status	Retained, Last Will “offline”	20 B
<tower>/heartbeat	Every 3 s: FPS, uplink	111 B
<tower>/detection	QoS 1: bearing, JPEG	70–90 KB
<tower>/weather	NB-IoT, short keys	58 B
server/fire	Retained fire position	155 B

Wildfire IoT04 / 10

Computer vision

Data and model

FIgLib

525

fire sequences from HPWREN tower cameras, one frame per minute, 40 minutes before and after ignition

Test set today

Pyronear

YOLO11s

tower imagery and an open smoke detector, Apache-2.0, trained for early smoke

Baseline today

D-Fire

20k+

labelled smoke and flame images in YOLO format

Our training · next

NASA FIRMS

Years

of satellite hotspots for Kazakhstan, clustered with DBSCAN into a fire-risk map

Risk map · next

1024 pxinput size

19–29 msper frame on a laptop GPU

≥ 0.20confidence per frame

3 framesin a row before an alarm

Wildfire IoT05 / 10

Results

Detection on real fires

Fire	First hit	Alarm	False before
2019 · bh-w	+3	+7	0
2019 · lp-s	+5	+7	0
2020 · om-n	+5	+7	0
2020 · rm-w	–	missed	0
2020 · pi-s	+4	+6	0
2021 · wc-e	+0	+5	0
2023 · pi-s	+0	+6	0
2024 · ws-s	+1	+3	1

Minutes after ignition, HPWREN FIgLib sequences.

7/8fires detected

6 minmedian to alarm

1false alarm in 313 min before ignition

Wildfire IoT06 / 10

Working demo

Live pipeline

Operator console: two towers, bearing rays and the triangulated fire on a satellite map

video→edge node→MQTT→backend→map + Telegram

W

WILDFIRE

bot

🔥 Smoke detected · TOWER-01
Confidence 0.55 · bearing 111.0°

14:59:31

🔥 Smoke detected · TOWER-02
Confidence 0.76 · bearing 228.9°

14:59:36

📍 Fire located by triangulation
43.07520, 76.93027 · 20.0 km / 13.8 km

14:59:36

Wildfire IoT07 / 10

Localisation

Locating the fire

bearing = heading + atan((x − W/2) / f)
f = (W/2) / tan(FOV/2)

1 towergives a ray. Range along it needs terrain: ray × SRTM elevation model (next)

2 towersrays cross at the fire. Today: 20.0 km and 13.8 km, crossing at 118°

Lasernot used: it reaches a few km and smoke is not a solid target

DJI Mini 4 Pro · final stage

Fly to the fix, confirm smoke from close range, and measure localisation error against the drone's GPS.

Geometry is a staged scenario: HPWREN footage replayed on two virtual towers near Almaty.

Wildfire IoT08 / 10

Interactive

Place a fire

Click anywhere on the map. Each tower decides if it can see the smoke, finds it in its frame and turns the pixel into a bearing.

Satellite map of the Ile-Alatau foothills south of Almaty

Imagery: Esri, Maxar, Earthstar Geographics

Camera

Sensor

Telegram: offline

No rangefinder. A laser reaches a few kilometres and smoke is not a solid target. Distance comes from a second tower, or from terrain (DEM) for a single tower.

Wildfire IoT09 / 10

Plan by weeks

Done and next

Weeks 1–7

Review and datasets
YOLO baseline, evaluation
IoT architecture, MQTT

Weeks 8–10

Temporal filter
Edge node
Sequence model
Pi / Jetson

Weeks 11–13

Map and alerts
Terrain range
Risk map, fuzzy
Drone check

Weeks 14–15

Experiments
Report, defence

Thank you

Wildfire IoT10 / 10