Wildfire & environmental monitoring from cameras

Wildfires are often reported by passers-by or spotted from fire watchtowers, so minutes to hours are lost before the first crew is dispatched. AI smoke detection runs continuously on images from tower-mounted or mountain-top cameras, flags a plume at long range, and sends an alert with a picture series to a human dispatcher who confirms or dismisses it. The same pipeline pattern also serves other environmental event detection (illegal dumping, industrial smoke, river pollution, flooding).

Problem & value

• Today: detection relies on emergency calls, patrols, satellite hotspots (revisit delay, coarse resolution) and, in some regions, human-staffed lookout towers or classic optical smoke-detection systems with hand-tuned image rules and many false alarms (haze, dust, clouds, fog, steam).
• AI value: 24/7 attention at constant quality, earlier alerts while a fire is still small and suppression is cheap, triangulated position from several cameras, and one operator watching a whole region. Time-to-detection is the key figure; the evidence base is mostly vendor and agency pilots, not independent trials, so measure it yourself before committing.
• Climate pressure: the European Forest Fire Information System (EFFIS) reports burnt area each season across the EU, with the Mediterranean worst hit and fire regimes moving north into central Europe, which includes Germany's pine forests in Brandenburg and Saxony [1].

Variants

• Forest-fire early warning (core): fixed PTZ towers, 10-30 km range, state forestry or fire services.
• Utility corridor monitoring: cameras on pylons/substations to catch ignitions near power lines (see VIS-003).
• Industrial & waste sites: smouldering fires in recycling yards and waste bunkers (thermal cameras).
• Drone & rapid-response: drones with thermal payloads confirm alerts and map fire fronts.
• Environmental monitoring: illegal dumping at hot-spots, river foam/oil sheens, algae blooms, snow/glacier and water-level time-lapse, dust and emission plumes at quarries and plants.
• Private/community: low-cost cameras for rural properties or municipality-run networks.

How it works

1. camera (visible/NIR/thermal, PTZ sweep or fixed)
2. frame/crop extraction on edge box
3. detector / classifier (smoke, flame, "other") + temporal filter (motion, growth over frames)
4. context filters (known sources: chimneys, fog banks, dust, sun glare; wind & time of day)
5. multi-camera triangulation
6. alert package (image series, bearing, map position)
7. human verification in control room (accept / reject)
8. dispatch (CAD, SMS, app)
9. feedback labels
10. retraining

• Detection: convolutional or transformer detectors (YOLO-family, RT-DETR) on high-resolution crops, often tiled because early smoke covers few pixels; temporal models reduce false alarms by requiring a plume to persist and grow. Vision-language models help as a second-opinion filter that explains the alert ("white plume rising from ridge, wind-blown") and rejects clouds and steam.
• Edge vs. cloud: remote towers have limited bandwidth, so inference runs on-site (GPU/NPU edge box) and only alerts and snippets are sent; models are updated centrally.
• Human-in-the-loop: a trained operator always confirms before crews are dispatched; operators' rejections are the most valuable training data.

Data preparation

The system needs live camera streams plus calibration data, and for tuning also a library of labelled smoke, fog and dust images. Camera streams usually come from tower installations, VMS servers or drone ground stations; site metadata (position, bearing, elevation) sits in GIS or spreadsheets; weather and fire-danger indices come from public services such as DWD. Weather data and satellite hotspots are accessible via APIs and need little work.

Typical gaps: no archived footage of real events, so few positive examples; look-alikes (fog, dust, steam, cloud) without labels; cameras with unknown orientation or drift; poor night and thermal coverage; unclear usage rights for recordings.

Preparation steps: inventory cameras and streams, record exact geo-calibration, collect and label 500-2,000 false-alarm and true-event frames with local rangers (see PREP-018 and PREP-020), set up a feedback loop for dispatcher corrections (PREP-021), and mask private property and passers-by in the image. Tools: CVAT or Label Studio, ffmpeg, a VMS with RTSP export. Roughly 30-40 percent of the effort goes into data and calibration; see GUIDE-003.

Implementation paths

Which path for whom

• Enterprise (utilities, forest owners): Path A, because mature turnkey smoke-detection services with cameras, models and alerting exist and the process is a commodity monitoring function.
• Public (forest services, municipalities): Path C on sovereign EU or German infrastructure, because procurement rules and sovereignty requirements favour operating open components on own or national cloud capacity.
• Developer: Path C, because open detectors (YOLO-family, fine-tuned on public smoke datasets) run on edge GPUs and allow integration with CAD and weather data.

Path A - Buy

• Fire-detection camera networks: Pano AI (US; ultra-HD towers, human review service), FireWatch by IQ wireless (Germany, EU; long-established optical system deployed in German forests), Pyronear (France, EU; non-profit-origin, low-cost cameras, open-source software), Dryad Networks (Germany, EU; solar gas sensors for very early smouldering-fire detection - a sensor complement, not cameras). In the US the university-run ALERTCalifornia network shows the public-camera model [2][3][4].
• Camera and VMS vendors (Axis, Hanwha, Teledyne FLIR, Bosch, Hikvision-free options for sensitive sites) offer thermal and AI analytics for industrial fire and hot-spot detection.
• Choose A when: you need tested hardware, mast installation, radio/solar power and a managed alerting service; building this yourself is mostly integration and maintenance work.

Path B - Build on cloud APIs

• Camera ingest: RTSP/ONVIF to a stream service (AWS Kinesis Video, Azure, GCP; EU regions).
• Models: Google Gemini or other multimodal APIs for second-opinion triage of alert snapshots; Amazon Rekognition Custom Labels / Azure Custom Vision / Vertex AI for a small custom detector.
• Orchestration: Node-RED/n8n or serverless functions to route alerts into dispatch tools.
• Choose B when: you already have cameras and want a pilot or secondary AI layer without hardware changes.

Path C - Self-host (open source)

• Pyronear (pyro-engine, pyro-vision, pyro-api, with a public smoke dataset on Hugging Face) is the most complete open stack for tower cameras [4].
• Detectors: Ultralytics YOLO (check AGPL/commercial licence), RT-DETR, RF-DETR; tracking via ByteTrack; the public Pyronear smoke dataset for pretraining [5].
• Serving: NVIDIA Jetson Orin or similar edge devices at the tower (HW1), a central GPU server (HW2) for review, Frigate or Viseron for generic NVR integration, MQTT for alerts.
• Effort: needs ML and field-hardware skills (power, radio links, lens/housing, calibration).

Path D - Custom training

Fine-tuning is almost always required for a new region: vegetation, haze, terrain and camera type change what smoke looks like. Collect several thousand labelled smoke sequences plus a large pool of hard negatives (clouds, fog, dust, burn piles, industrial steam) from your own sites; synthetic smoke compositing helps with rare events. Keep a held-out, per-site test set.

State of the art (2026)

• Operational networks in California, Australia, Portugal, Spain, France and Germany use AI detection with human confirmation; reported performance depends heavily on site, season and weather, so compare vendors on your own recorded footage.
• Techniques: high-resolution tiling, temporal/sequence models, multi-camera triangulation, fusion with weather, satellite hotspots (Copernicus, NASA FIRMS) and gas sensors; open-vocabulary detectors and VLMs for zero-shot environmental events.
• Limits and failure modes: false positives from fog, low sun, dust, cooking and industrial steam; night detection depends on IR/thermal and flame glow; smoke hidden behind terrain; heat shimmer; camera drift needing recalibration; very small distant plumes at the edge of range. Missed detections matter more than false alarms, so thresholds are set operationally.

Shared resources & combinations

• Same stack (object-detection + tracking + edge-runtime): VIS-010 intrusion detection, VIS-012 traffic analytics, VIS-032 camera-trap monitoring, VIS-060 turbine bird-collision cameras; one edge-box fleet and one NVR/MQTT backbone can host several models.
• Complementary: VIS-021 and VIS-042 (satellite change detection and plume detection), VIS-058 (post-fire damage mapping), VIS-031 (natural-language search over camera archives), PRED-027 (flood forecasting), VIS-003 (drone inspection of lines near vegetation).

Effort, cost & KPIs

• Path A: weeks to a few months for tower siting, permits, power and connectivity; costs are dominated by hardware, masts, communications and service fees, not the AI model.
• Path B: 1-3 months for a pilot on existing cameras.
• Path C/D: 6-12+ months to reach reliable operation across seasons (needs at least one full fire season of data).
• KPIs: detection recall at fixed range, time from ignition to alert, false alarms per camera per day, share of alerts confirmed by operators, area covered per camera, operator time per alert, system uptime (power, link), time from alert to first crew on scene.

People & change

Start with management and the department lead (forest or operations) to fix the scope and alerting chain, and involve IT and information security early because the cameras, radio links and alert channels are critical infrastructure. Data owners (land owners, camera operators) must agree to use of the streams. The dispatchers and rangers who receive alerts are the real end users; let them tune thresholds in a pilot on two or three towers in one season before wider rollout.

Works council: usually not triggered, as the cameras look at forest and not at staff. It depends where cameras or drones can capture employees, for example at a plant or depot; then the objective suitability for monitoring performance or behaviour triggers co-determination under §87(1) no. 6 BetrVG, and the council must in any case be informed under §90(1) no. 3 when work processes are planned with AI. Public bodies involve the staff council (Personalrat).

Data protection: involve the DPO, check whether people can be identified, use masking and short retention, and sign an AVV with any cloud vendor. A DPIA is likely only if public areas are covered at scale. For AI literacy, support dispatchers with short training on false alarms and limits (Art. 4 AI Act). Communicate that humans always confirm alerts before dispatch. Guides: GUIDE-008, GUIDE-009, GUIDE-005.

Risks, regulation & compliance

• EU AI Act (Regulation 2024/1689): a typical forest fire detector is minimal risk; it is not prohibited under Art. 5 and not an Annex III system. It could become high-risk only when it is a safety component in the management and operation of critical infrastructure (Annex III point 2) or integrated with emergency dispatch decisions (Annex III point 5(d), emergency call and dispatch triage) - assess case by case. Art. 4 AI literacy applies to operators [6].
• GDPR: long-range cameras can incidentally capture people, vehicles, homes and licence plates. Apply privacy masks, limit pixel density on residential areas, restrict retention, document a legitimate-interest or public-task basis and a DPIA where public areas or neighbourhoods are covered. Environmental dumping cameras that identify offenders enter surveillance territory (see VIS-010).
• Other rules: national forest and fire-protection law, radio and mast permits, nature- protection constraints, flight rules for drones (EU 2019/947), and CER/NIS2 duties for critical operators.
• Operational risks: alert fatigue (tune thresholds, require human review), over-reliance in poor visibility, single points of failure (power, link) - combine with sensors, patrols and satellite data.

Getting started

1. Define the region, fire-danger season and required alert time; map line-of-sight coverage (viewshed analysis).
2. Record 2-4 weeks of footage from candidate sites, including hazy and foggy days.
3. Run a vendor bake-off (FireWatch, Pano AI, Pyronear) on that footage; measure recall and false alarms.
4. Pilot 2-3 overlapping cameras with operator confirmation and a documented alert-to-dispatch process.
5. Complete the DPIA, set retention/masking rules, then scale with a feedback loop for retraining.

Sources

1. EFFIS - European Forest Fire Information System (Copernicus) - https://forest-fire.emergency.copernicus.eu/
2. Pano AI - wildfire detection - https://www.pano.ai/
3. ALERTCalifornia (UC San Diego) - https://alertcalifornia.org/
4. Pyronear - open-source wildfire detection - https://github.com/pyronear
5. Pyronear datasets on Hugging Face - https://huggingface.co/pyronear
6. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Dryad Networks - Silvanet wildfire sensors - https://www.dryad.net/
8. NASA FIRMS - Fire Information for Resource Management System - https://firms.modaps.eosdis.nasa.gov/