Ground-Based Fire Detection Camera Networks: Global Experience and 2026 Limits

April 20, 2026

Posted in Blog

Ground-Based Fire Detection Camera Networks: Global Experience and 2026 Limits

On 17 August 2024 at 14:23 Pacific time, the AlertWildfire camera Snow Mountain 2 (2,100 m above Cloverdale, Sonoma County) caught a thin smoke column on azimuth 218°. CAL FIRE crews arrived 22 minutes later. The Point Fire was held at 4 acres — a number that would read 100 to 1,000 acres without the camera. This review maps global ground-based detection: AlertWildfire’s 1,100 cameras, Mediterranean CCTV, the Australian hybrid, and the cost arithmetic that keeps WildFiresUA satellite-first across 600,000 km² in 2026.

Viewshed geometry: visibility and elevation

The first determinant of camera utility is its viewshed: the set of terrain points visible from a given observation site, accounting for relief and the curvature of the Earth. A camera at 100 m on flat terrain theoretically sees out to 35 km, but practically only to 5-10 km due to atmospheric refraction and visibility. A camera on a 2,000 m mountain peak sees out to 160 km, but practically to 50-80 km for a 0.5° angular smoke-column size.

Allison et al. (2016) review for the journal Sensors and analyse in detail the trade-off between installation height, camera count, and area-coverage completeness (Allison et al., 2016, Sensors). Their key conclusion: 90% area coverage in hilly terrain requires 1 camera per 200-500 km². Flat terrain without high installation points (typical for Ukraine, Poland, Lithuania) requires 1 camera per 50-150 km².

This means national coverage of Ukraine (60 million ha = 600,000 km²) would require 4,000-12,000 cameras with installation on 30-60 m masts. For comparison, AlertWildfire covers about 200,000 km² (California plus parts of adjacent states) with 1,100 cameras, or 1 camera per ~180 km² in mountainous terrain.

Toulouse-Inok et al. (2018) proposed an optimisation algorithm for camera-node placement that accounts for risk-zone priority (Toulouse-Inok et al., 2018, IEEE TGRS). Priority high-risk zones (residential wildland-urban interface) can require densities of 1 camera per 20-50 km², while low-risk zones can accept 1 camera per 1,000+ km².

AlertWildfire: the American benchmark

AlertWildfire is the largest ground-based camera network in the world. Founded at the University of Nevada, Reno (UNR) in 2014, it has grown to over 1,100 cameras across California, Nevada, Oregon, Idaho, Washington, Colorado, Utah, and Arizona. Partners include UC San Diego, the University of Oregon, NV Energy, Pacific Gas & Electric, and Southern California Edison. Public access is at alertwildfire.org.

Technical architecture:

Sony FCB-EV9520 cameras or analogues with 30x optical zoom, 360° PTZ, and NIR night mode.
Solar panels of 200-400 W and a 200-400 Ah battery for autonomous operation in remote locations.
Connectivity through cellular (LTE), microwave, or satellite (Starlink) depending on availability.
RTSP streaming to the central UNR server with 1-3 second latency.
Archive retention: 7 days full, 30 days compressed.

Workflow: an operator at the dispatch centre monitors a mosaic of 12-24 cameras simultaneously. On a smoke suspicion, the operator zooms in, cross-checks with other cameras within the viewshed (cross-verification), and if confirmed dispatches the fire service. Time from first trigger to dispatch is 5-15 minutes depending on operator load.

ALERTCalifornia: the AI extension

ALERTCalifornia is a separate UC San Diego project that adds an AI layer to AlertWildfire. Launched in 2023, ALERTCalifornia analyses camera video feeds in real time using a deep neural network (CNN with residual blocks, trained on thousands of smoke and false-positive examples) (ALERTCalifornia).

Govil, Welch, Ball, Pennypacker (2020) published the first results: AI pre-screening detects smoke an average of 7-10 minutes earlier than human operators, at a false positive rate near 7% (Govil et al., 2020, Remote Sensing). The increased false-positive rate is offset by AI filtering out the obvious negatives (clouds, fog, glare), leaving only suspicious cases for the human.

AI architecture: as of 2024 ALERTCalifornia uses a model with a ResNet-152 backbone plus a temporal attention module that analyses sequences of 6-12 frames. The training dataset contains over 10 million labelled frames from the 1,100+ cameras.

Critical hyperparameters:

Classification threshold of 0.75 (a compromise between false positive and false negative).
Temporal aggregation window of 60 seconds.
Number of consecutive positive frames required for an alarm: 4 of 6.

Cleartrip et al. (2022) published a comparison of several CNN architectures (ResNet, EfficientNet, Vision Transformer) on a shared fire dataset. ResNet-152 won on F1, while ViT-B/16 won on inference speed (Cleartrip et al., 2022, Fire).

Europe: Mediterranean CCTV

Europe lacks a single continent-wide network. Deployment proceeds at the national and regional levels.

Greece has one of the most developed forest-monitoring CCTV networks, with about 800 cameras in priority regions (Attica, Peloponnese, Crete). Coordination is through the Hellenic Forestry Department and Civil Protection (Pyrosvestiki). Most cameras are integrated with radio weather stations for joint visibility and temperature monitoring.

Spain deployed PLAN INFOCA in Andalusia: 162 cameras cover about 4.2 million ha. Catalonia, Galicia, and Castilla-La Mancha each operate separate networks of 50-200 cameras. Lozano et al. (2018) describe the PLAN INFOCA technical implementation and compare it with American systems (Lozano et al., 2018, Forestry).

Portugal, after the catastrophic 2017 fires (114 fatalities), invested heavily in CCTV: about 600 cameras in northern and central regions. Coordination is through ANEPC (the National Authority for Emergency and Civil Protection).

Italy takes a fragmented approach: each region has its own network. Sicily, Sardinia, and Calabria each have 50-150 cameras. National coordination runs through DPC (Dipartimento della Protezione Civile).

Croatia, Slovenia, Montenegro operate small coastal networks of 30-80 cameras focused on tourist regions.

AI camera vendors: the commercial segment

A separate market segment is commercial AI camera detectors. Major players include American and European category leaders that offer end-to-end “camera + AI + dashboard” solutions on a SaaS model.

Typical commercial vendor configuration:

Purchase or installation on existing masts of 360° PTZ cameras with 30-50x zoom.
Edge-AI deployment on the device or in the vendor data centre.
SLA: trigger-to-push-notification latency of 60-120 seconds.
Pricing model: 12,000-25,000 USD per camera per year (including hardware, AI, and operations centre).

Who pays: insurance companies (risk modelling), utility power networks (Pacific Gas & Electric, Southern California Edison invest in networks to prevent power-line ignitions, a consequence of the financial shock from Camp Fire 2018, which bankrupted PG&E), federal forestry agencies (USFS, USDA), and regional forestry departments.

Ahmad, Bao, Khan (2017) provide a review of commercial AI smoke-detection models in IEEE Access and discuss the trade-offs between accuracy and inference speed (Ahmad et al., 2017, IEEE Access).

The technical task: smoke vs steam vs cloud

Optical smoke detection is non-trivial. Smoke visually resembles:

Low-level clouds (cumulus, stratus).
Valley fog in morning hours.
Steam from water bodies (warm lakes on a cold day).
Dust from agricultural work (combine, ploughing).
Industrial emissions (thermal power plants, metallurgy, cement plants).
Vapours from asphalt-concrete plants.

Classical computer-vision methods (HSV analysis, Gabor texture analysis, optical flow) work on simple scenes but produce high false-positive rates in natural conditions.

Modern CNN approaches add several key features:

Spatio-temporal consistency: smoke has a characteristic vertical growth speed (1-5 m/s) that differs from clouds (slower) and dust (faster, but without a vertical component).
Texture analysis: smoke typically shows a fractal structure with a characteristic edge-element distribution.
Spectral features: for cameras with an NIR channel, smoke has lower NIR reflectance than clouds.
Contextual features: source position (smoke rises from a discrete point, while a cloud occupies a wider area) and presence of fire-prone weather conditions (modelled with meteorological input).

Frizzi, Bouchouicha, Ginoux (2016) proposed a CNN architecture for joint smoke and flame detection in video feeds with F1 = 0.87 on the test dataset (Frizzi et al., 2016, IECON Proceedings). Tao et al. (2019) extended the approach to multi-task learning, with the model jointly performing classification and smoke-area segmentation (Tao et al., 2019, Fire Safety Journal).

Camera system limits

Night operation remains a fundamental constraint. Without active illumination (infrared or thermal) PTZ cameras with optical zoom provide no information in darkness. Thermal imagers add 3-10 times to cost, sharply limiting scaling. Many systems (including AlertWildfire) operate primarily during the day and rely on satellite alarms at night.

Fog and high humidity fully block optical detection at distances above 1-3 km. In the Carpathians and Polissia of Ukraine this means autumn and winter morning detection can be inactive 30-50% of the time.

Maintenance: a camera mast requires power (solar panels, batteries), protection from rodents and vandalism, and periodic optical cleaning (pollen, spider webs, dirt). Typical maintenance frequency is 2-4 visits per camera per year. For a 1,000-camera network this means 2,000-4,000 field crew visits per year.

Connectivity: cellular can fail in remote locations. Microwave is weather-sensitive. Satellite (Starlink) is expensive but became an acceptable solution after 2022.

Cost: deployment of one camera site costs 25,000-80,000 USD (camera, 30-60 m mast, power, connectivity, installation) plus 5,000-15,000 USD per year in maintenance. A 5,000-camera national network would cost 125-400 million USD in capital and 25-75 million USD in annual operating expenses.

The Australian experience

Australia has one of the most demanding fire environments in the world: eucalyptus forests with high burn intensity, frequent atmospheric catastrophes (pyrocumulus, fire tornadoes), and dispersed populations in the wildland-urban interface.

The Bushfire Cooperative Research Centre (now Bushfire and Natural Hazards CRC) ran a series of camera-network studies between 2010 and 2020. The key result: for Australian terrain, satellite detection is 70-80% cheaper and more efficient over most of the territory, while camera networks are economically justified only within the wildland-urban interface and around critical infrastructure (Australia has no nuclear plants, but the same logic applies to oil-refinery and chemical sites).

Regional VIC, NSW, and ACT have point camera deployments in the WUI background, coordinated through local agencies CFA (Country Fire Authority Victoria) and RFS (NSW Rural Fire Service). The 2024 standard is a hybrid architecture combining cameras, the Himawari-9 geostationary, and VIIRS polar.

Williams et al. (2017) analyse in detail the economic effectiveness of Australian camera networks. Conclusion: ROI is positive for zones with building density above 10 houses per km² and negative for remote forest tracts (Williams et al., 2017, International Journal of Wildland Fire).

Mounting infrastructure: masts and towers

Camera-network quality depends on the availability of high installation points.

The United States uses a combination of:

Former forestry agency lookout towers (USFS holds over 5,000 historical positions, of which about 800 remain active).
AT&T, Verizon, and T-Mobile telecommunications masts in vendor partnerships.
High-voltage power lines. PG&E deploys cameras on its own high-voltage lines.
Dedicated 30-60 m masts purpose-built for the fire network.

Europe relies primarily on existing telecommunications masts plus dedicated forestry towers. Greece and Portugal use serpentine mountain roads with small towers on each pass.

Ukraine has a dispersed communications infrastructure: Kyivstar, Vodafone, and lifecell hold tens of thousands of masts across the country. The challenge for fire monitoring is that mast placement is optimised for population coverage, not forest tracts.

Examples of operational metrics

Network	Cameras	Coverage km²	Alarm time	FP rate
AlertWildfire	~1,100	~200,000	5-15 min	3-8% (with operator)
ALERTCalifornia AI	~1,100	~200,000	3-8 min	7-12% (AI alone)
PLAN INFOCA (Andalusia)	162	~42,000	10-20 min	5-10%
Greek forest monitoring	~800	~100,000	8-15 min	8-15%
Portuguese ANEPC	~600	~80,000	7-15 min	6-12%
Australian VIC/NSW regional	~250	~40,000 (WUI)	5-12 min	5-10%

Future directions: thermal imagers, multispectral cameras, edge AI

Thermal imagers with uncooled microbolometric sensors are becoming more accessible. A FLIR Boson 640 thermal camera now costs about 6,000-10,000 USD instead of 30,000+ USD ten years ago. This makes night detection realistic for new deployments. Bushfire CRC estimates that 30-50% of new cameras in California will have thermal imagers by 2030.

Multispectral cameras with RGB + NIR + SWIR channels form a separate class. SWIR (short-wave infrared, 1-3 micrometres) discriminates smoke from clouds well, since water molecules in clouds absorb strongly in SWIR while smoke aerosols absorb less. Camera vendors in the “satellite-grade for ground use” category offer such solutions for critical infrastructure at 15,000-40,000 USD per unit.

Edge AI: NVIDIA Jetson Orin and analogues allow CNN inference directly at the camera node without sending raw video to the data centre. This reduces bandwidth by 100-500 times and latency by 1-3 seconds. The 2025-2026 standard is edge AI in new AlertWildfire and PLAN INFOCA deployments.

Federated learning enables continuous model retraining on new false-positive and miss examples. This allows the model to adapt to seasonal change (in autumn, falling yellow-brown leaves trigger smoke classifications, and these cases train the model to ignore them).

Ukraine: why WildFiresUA does not use cameras as its core layer

Ukraine has several structural characteristics that limit the feasibility of a camera network as a core layer.

Geography: mostly flat terrain with few high points. The Carpathians, Crimean mountains, and Donetsk Ridge cover only 5% of the territory. Coverage of the flat forest tracts of Polissia, Slobozhanshchyna, and the steppe zone therefore requires many short masts (30-60 m) rather than a few high mountain positions.

Territory size: 60+ million ha of relevant forest and steppe zones. Even 1 camera per 200 km² requires 3,000+ cameras. CAPEX 75-240 million USD, OPEX 15-45 million USD per year. For a country at war without significant external funding, this is economically unrealistic.

War: most forest tracts of Polissia, Slobozhanshchyna, Donetsk and Luhansk regions contain mined territories, defensive lines, and restricted access for civilian personnel. Mast deployment and maintenance is impossible across many priority zones. Existing telecommunications masts are partially damaged by shelling.

Electronic warfare: cellular streaming in front-line zones is unreliable due to electronic warfare, which limits connectivity reliability.

Within this context, WildFiresUA, the national fire service from the YourAirTest team in partnership with Oles Honchar Dnipro National University (DNU) and EcoCity, is built on a satellite-first principle. The active-fire map at partner.yourairtest.com/map integrates VIIRS, MODIS, and Sentinel-3 SLSTR — globally available streams that do not depend on ground infrastructure inside Ukraine.

A camera network is considered as a future second layer, primarily in two zones: the Carpathians (where mountain peaks provide natural high points and seasonal fire risk) and the area adjacent to the Chornobyl Exclusion Zone (where the radiation component requires especially early detection). These pilot deployments are planned with international partners and are under team consideration on a 2027-2029 horizon.

Summary

Ground-based camera networks are a powerful detection layer in regions with compact territory, high infrastructure density, and adequate funding. AlertWildfire demonstrates that 1,100 cameras can deliver detection for 90%+ of fires in California with mean alarm times of 5-15 minutes. European systems (Greece, Spain, Portugal) show that a CCTV approach also works in Mediterranean climates. The ALERTCalifornia AI extension reduces alarm time to 3-8 minutes at the cost of a higher false-positive rate.

Camera networks have critical limits, however: high deployment and maintenance cost, limited night operation without thermal imagers, sensitivity to fog and humidity, and a need for high installation points. For countries with large territory, flat terrain, and constrained budgets (Ukraine included), a camera network as the core layer is economically unrealistic. The satellite layer (VIIRS, Sentinel-3 SLSTR, MTG FCI) provides acceptable national coverage at sharply lower cost by relying on globally available state infrastructure.

WildFiresUA in 2026 operates as a satellite service for Ukraine, with possible future addition of camera nodes in priority zones (Carpathians, Chornobyl) in partnership with international funds. This approach aligns with the best practice of Brazil, Canada, and Australia — countries with large territory where camera networks complement rather than replace satellite detection.

Ukrainian startup ecosystem: follow TechUkraine and AIN.ua — the two leading outlets covering Ukrainian deep tech, climate tech, and environmental startups.

What to do today

Check the YourAirTest air quality map for your city — recent PM2.5 readings.
If the topic resonates, share with colleagues; we monitor referrals via Google Search Console.
To contribute data (sensor measurements, regional models) reach out via the contact form.