Early Wildfire Detection in Brazil: 95% Recall and Under-5-Minute Alerts

Brazil’s forests and farmland lose millions of acres annually to wildfire, with the Pantanal recording a 462% surge in fire events in one year alone. Omdena partnered with Sintecsys, a Brazilian fire monitoring company, to build a CNN-based smoke detection system covering 8.7 million acres across 50 tower-mounted cameras. The 8-week project delivered models achieving 95–97% recall, an 8,750-image labeled dataset, and a reduction in average detection time from 40 minutes to under five minutes.

Watch the Webinar

See how AI can help detect wildfires earlier and support faster, more informed response.

In this webinar, we explore the computer vision approach, AI models, and key challenges behind the project, along with its results and practical applications for earlier wildfire detection and faster response.

The Problem: Wildfire Detection at Scale Across Brazil’s Forests and Farmland

Wildfires are a recurring crisis in Brazil. In 2020 alone, over 8,000 fires were recorded in the Pantanal, marking a 462% increase over the prior year. Globally, wildfires contribute one-third of total CO₂ emissions, damaging ecosystems, harming respiratory health for communities downwind, and generating significant financial losses for agricultural companies.

The client monitored 8.7 million acres of Brazilian forest and farmland using 360-degree cameras across seven states and four biomes. Operators reviewed live feeds around the clock, manually verifying each alert before dispatching firefighters. By 2019, average detection time stood at 40 minutes, long enough for a small fire to become uncontrollable.

The Partnership: Omdena and Sintecsys

Founded in 2016, the client operates a real-time fire outbreak detection and land management platform serving agricultural companies, forest owners, and municipal authorities across Brazil. Its monitoring system uses 360-degree cameras on 50 communication towers and transmits live imagery to a central facility, where operators watch for the first signs of smoke around the clock.

Omdena partnered with the client to build an AI wildfire detection system using computer vision and deep learning. Over eight weeks, 47 collaborators from 22 countries worked alongside the client’s internal AI team to design, train, and evaluate models capable of detecting smoke in daytime imagery with high accuracy and low false positive rates.

Why Early Wildfire Detection Is Hard: Visual Noise, Missing Data, and Changing Light

Visual ambiguity at distance: Smoke in a camera feed is nearly indistinguishable from fog, low cloud cover, dust kicked up by vehicles or wind, and emissions from agricultural equipment. Any model trained to catch smoke aggressively will also flag these look-alikes; any model tuned to suppress false alarms risks missing the first sign of a real fire.

No existing labeled dataset: No open, labeled wildfire dataset existed matching the resolution, camera angle, and conditions of the client’s tower network. Every training image had to be extracted from the client’s own footage and manually annotated — a substantial pre-training workload before model development could begin.

Boundary lighting conditions: Sunrises, sunsets, and shifting weather patterns alter how smoke appears against the sky, creating conditions where even consistent image quality does not guarantee stable model performance. Separating genuine smoke from sky color changes at dawn and dusk required careful attention during both labeling and data augmentation.

Building the Detection System: Dataset Curation, CNN Architectures, and Parallel Experimentation

Step 1 — Data Collection

The team assembled a training corpus of approximately 8,750 images from the client’s camera network, covering fire and non-fire scenes across different times of day. Gary Diana developed an algorithm to extract frames from video footage without generating duplicates, adding 1,150 frames at 1,280 × 720 pixels to the initial 7,600-image base.

Step 2 — Labeling and Curation

Labeling was handled by approximately 20 volunteers using Labelbox, which enforced role separation between labelers, reviewers, team managers, and administrators. This structure prevented contributors from overwriting one another’s work and enabled consistent quality control across a corpus mixing genuine smoke events with fog, dust, cloud, and glare.

To improve generalization, the team applied horizontal and vertical flips, image patches of varying sizes, and upsampling for lower-resolution frames. Label smoothing helped models distinguish genuine smoke from boiler emissions and lens flare.

Step 3 — Model Architecture Exploration

Three parallel tracks ran simultaneously: MobileNet architectures for edge deployment on lower-powered hardware; semantic segmentation to localize smoke at the pixel level; and CNN classifiers ranging from simple binary models to deep feature extraction networks. CNN-based classifiers produced the highest recall on the client’s held-out test images and were selected as the foundation for the deployable system.

What Was Delivered: A Deployable Daytime Smoke Detection Model and Training Pipeline

CNN-based smoke detection model: A trained convolutional neural network that analyzed daytime camera images in real time, flagging frames containing smoke and suppressing alerts for common look-alikes. Initially deployed as a prototype with one client, the model integrated with the existing monitoring platform without requiring new hardware; full rollout to all customers was planned for end-2020.

Labeled training dataset of 8,750 images: A curated corpus drawn from the client’s own camera network, covering fire and non-fire scenes across multiple biomes and lighting conditions. The dataset included smoke, fog, dust, cloud, and glare examples, enabling the model to learn the differences between genuine outbreaks and visual noise at scale.

Full open-source training pipeline: The complete preprocessing, augmentation, and training workflow was delivered as open-source code through OmdenaLore. The client retained the ability to retrain on new footage and extend the system to nighttime detection in a planned second phase.

Impact: Faster Detection and Fewer False Alarms Across 8.7 Million Acres

The model achieved 95–97% recall on held-out daytime test images, identifying nearly every genuine fire event in the evaluation set. Average detection time dropped from 40 minutes to under five minutes, and the client reported a 90% reduction in crop and tree losses across monitored territories.

Before the AI layer, operators reviewed every alert including false triggers from fog, clouds, and dust. The false positive rate varied between 10% and 33% depending on camera location — forest towers produced fewer spurious alerts than cameras near agricultural equipment and road traffic. With the model handling initial classification, operators focused on events already screened as likely genuine.

Fewer uncontrolled burns meant less CO₂ released from Brazil’s biomes. Faster detection also reduced smoke exposure for farming communities near monitored land, where respiratory illness linked to wildfire smoke is a documented public health concern.

Benefits for Clients: Automated Monitoring Without New Infrastructure

Automated first-pass filtering on existing camera streams: The model processed every incoming camera frame before it reached a human screen, eliminating manual review of fog, dust, and cloud events. Operators received fewer alerts and focused on events the model had already flagged as likely genuine.

No specialist AI skills required to operate: The client’s monitoring staff continued their established review workflow unchanged. The AI layer slotted in as a pre-filter, with no retraining of personnel, no changes to team structure, and no new tools to learn.

Open-source training pipeline via OmdenaLore: The full dataset, model weights, and training code were delivered as open-source assets through OmdenaLore. The client retained the ability to retrain on new data and scale to additional camera networks without licensing constraints or vendor dependency.

What Building a Wildfire Detection System for Brazilian Forests Taught Us

A model optimized for recall outperforms a precision-focused model in fire monitoring contexts

The team prioritized recall and accepted a 10–33% false positive rate as a deliberate trade-off. A missed fire carried far greater cost than a false alarm cleared by an operator in seconds — a priority that shaped every architecture decision and confidence threshold tested.

Label quality determines model quality more than architecture choice in computer vision projects of this scale

Assembling and labeling approximately 8,750 images required more effort than training the models themselves. The Labelbox workflow and role separation made the corpus reliable — switching architectures cannot compensate for noise introduced at the annotation stage.

Defining a narrow first-phase scope is what makes a complex system deployable on time

Limiting Phase 1 to daytime smoke detection allowed the team to deliver within 8 weeks. Nighttime detection, satellite integration, and predictive risk mapping were real requirements — scoping them as later phases, rather than attempting all conditions in parallel, was what made the initial deployment achievable.

What’s Next: Nighttime Detection, Satellite Integration, and Predictive Risk Mapping

Nighttime flame detection (Phase 2): Flames at night appear as bright point sources against dark backgrounds — a fundamentally different visual signature from daytime smoke, requiring separate labeled data and adapted model architectures. Integrating day and nighttime detection will give the platform consistent capability around the clock.

Satellite imagery integration: Ground-level towers leave coverage gaps between their fields of view. Satellite feeds will supply continuous spatial coverage, identifying smoke plumes or thermal signatures in uncovered areas and routing alerts into the same monitoring dashboard operators already use.

Predictive risk mapping: More than 90% of Brazilian wildfires are linked to human activity. The client plans to develop AI tools that identify elevated-risk areas before smoke appears, using historical fire records, land-use patterns, and human activity data to concentrate monitoring where fires are most likely to ignite.

Key Achievements: Detection Performance, Response Time, and a Reusable Data Asset

A Smoke Detection Model That Catches Nearly Every Fire Event

The model’s 95–97% recall meant nearly every fire event in the test set was identified at the earliest visible stage, giving firefighting teams time to intervene before flames spread — a response window the previous 40-minute baseline made impossible.

An 87% Reduction in Average Detection Time Across 8.7 Million Acres

The 35-minute reduction in alert time changed what a monitoring team can do. A fire caught within five minutes burns a fraction of the area lost at the 40-minute mark. The 90% reduction in crop and tree losses across monitored territories reflects that operational difference.

A Labeled Dataset Built from Production Camera Footage

Generic wildfire datasets do not reflect the specific camera angles, lighting conditions, and vegetation types of any one system. The 8,750-image corpus gives the client a foundation to extend for Phase 2 nighttime training without restarting data collection and annotation from scratch.

About the Project

Omdena partnered with Sintecsys, a Brazilian fire monitoring company, to build an AI wildfire detection system using computer vision and deep learning. Over eight weeks, 47 collaborators from 22 countries assembled and labeled 8,750 images, trained CNN-based smoke detection models, and delivered a system achieving 95–97% recall that reduced detection time from 40 minutes to under five minutes.