AI Applied: How to Stop Wildfires through Flame and Smoke Detection

The year 2019 delivered a stark warning about the devastating impact of uncontrolled fires. From the flames that devoured the Notre‑Dame cathedral in Paris and the National Museum in Brazil to widespread conflagrations in the Amazon forest fires and, later, the blazing landscapes of Australia, the world witnessed how quickly ecosystems and cultural heritage can be reduced to ash. Before discussing how ai powered wildfire detection can offer an early warning system, it is useful to understand how fires start in the wild and why they spread so rapidly.

In nature, blazes generally begin when lightning strikes dry vegetation. These natural fires account for a smaller share of incidents than human‑caused fires, but they often burn larger areas because they may go unnoticed for hours. Human activity contributes to man‑made fires through smoking, recreation, or agricultural practices, and these fires are more frequently detected early on. Regardless of origin, once a forest such as the Amazon ignites, flames can race across the landscape at speeds up to 23 kilometres per hour. Temperatures may soar to 800 °C (1,470 °F), destroying plant and animal life in hours, and vast plumes of CO₂ further aggravate global warming.

The destruction goes beyond the rural environment. In August 2019 a cold front met particulates from Amazonian and mid‑western fires and plunged São Paulo into darkness at mid‑afternoon. The city’s sky turned black at 3 p.m. on the 19th, creating an apocalyptic scene that left residents feeling as though they were living through a biblical plague.

 

In the midst of confusion and misinformation, scientific images clarified what was happening. A high‑resolution satellite photo posted by NASA illustrated how smoke traveled south‑east across Brazil, smothering cities in haze.

Throughout Brazil and elsewhere, fires have displaced thousands of people, claimed lives, destroyed property and signaled that such disasters will recur unless effective prevention and response strategies are implemented.

Harnessing AI to Stop Wildfires

Amid this backdrop, many people asked whether there was a way to combine community engagement and advanced technology to detect fires before they wreak havoc. According to the Brazilian wildfire detection company Sintecsys, the answer is an emphatic yes. The company installs cameras atop communication towers to watch for smoke or flames. When its system detects an anomaly, it transmits alerts to a monitoring centre that can trigger firefighting actions, saving lives and infrastructure. Although Sintecsys is not alone in striving to stop wildfires, its 50 towers (2019 data) cover a limited area. To extend the customer reach and scale their business model to thousands of cameras with accurate, rapid detection, Sintecsys partnered with Omdena.

Omdena is a global platform where organizations collaborate with a diverse AI community to build solutions for real problems in a faster and more effective way. Through this collaboration, Sintecsys sought to develop an ai powered wildfire detection system that could process live images and recognize smoke or flames with minimal delay.

How the Team Tackled the Problem

Scoping the challenge

Working together, Omdena and Sintecsys first agreed to address daytime images in their initial challenge. Daytime photos generally show smoke, whereas night‑time footage reveals actual flames. Sunsets and dawns, when smoke and fire coexist, represent boundary conditions that complicate image recognition. A second challenge would later incorporate night images to improve the detection pipeline.

Building and preparing the dataset

The team assembled a large corpus of footage and still images captured by different cameras, both with and without fire outbreaks. The original dataset comprised approximately 7,600 images with a resolution of 1,920 × 1,080 pixels, including daytime scenes without fires, daytime pictures with fires, and a smaller fraction (around 16 %) of night‑time images. To enlarge the corpus, Gary Diana wrote an algorithm that extracted additional frames from video footage while avoiding duplicates of identical landscapes. This contribution yielded about 1,150 more images at 1,280 × 720 pixels.

Labeling with Labelbox

Once the datasets were ready, roughly 20 volunteers set out to create labeled data. The team set up environments on Labelbox, a popular tool for computer vision projects that allows labeling, quality management and operation of a production training‑data pipeline. Team members ran tests and then began annotating the final datasets. I managed the task with invaluable support from Alyona Galyeva, who not only labeled data but also reviewed and coordinated the group’s work. She observed:

It always starts with a mess when a group of people collaborates on a labeling project. In our case, Labelbox saved us a lot of time and effort by not allowing multiple users to label the same data. On top of that, it made our lives easier by proposing 4 roles: Labeler, Reviewer, Team Manager, and Admin. So, nobody was able to mess with data sources, data formats, and, of course, the labels made by other people.

Labelbox’s role‑based workflow enabled the team to maintain consistency and avoid conflicts among annotations.

Labelbox interface for labeling, managing and reviewing labels

Building the models

From the outset, team members dived into state‑of‑the‑art research to identify techniques suited to wildfire detection. They split into sub‑teams, each exploring a different strategy such as MobileNet, semantic segmentation and various convolutional neural network (CNN) architectures ranging from simple to sophisticated. Danielle Paes Barretto described the atmosphere:

It was inspiring to see people eager to use their skills to stop wildfires and make an impact. I tried to help in all tasks; from labeling the data to building CNN models and testing them on our dataset. We also had frequent discussions which in my opinion is one of the greatest ways of learning. All in all, it was an amazing opportunity to learn and to use my knowledge for the good while meeting great people!

In parallel, the team applied different techniques to enhance performance: creating patches of varying sizes from the original images and training on those patches, augmenting data through horizontal and vertical flipping, and reducing noise in the imagery. By combining various model architectures with data‑centric tricks, the community strove to design a robust ai powered wildfire detection system.

Results

The initial models delivered promising metrics that underscored the viability of automated fire detection. In plain terms:

• High recall: The best solutions achieved recall rates between 95 % and 97 %, meaning they successfully identified the vast majority of actual fire outbreaks.
• Controlled false positives: These models maintained false positive rates between 20 % and 33 %, a tolerable level given the priority of catching real fires quickly.

Sintecsys was extremely pleased with these results. Nonetheless, the next challenge will further improve the models by incorporating night‑time images.

Learning process

As with every Omdena challenge, this project was a rich journey of hands‑on learning. There is no better teacher than direct engagement with a real‑world problem. Collaborator Iliana Vargas captured the spirit of the experience:

When I heard about Omdena, I did not think twice and I applied for the challenge. The experience I had in the project in general was very gratifying for me, not only in the technical part, but also in terms of being part of a community. We had an excellent team of professionals, but above all, we had people willing to contribute with their knowledge for a project that has a social benefit.

What’s next?

The logical next step is to continue improving the model to achieve even better results. This path involves building solid, cutting‑edge technology that not only strengthens Sintecsys’s position but also allows it to advance its business model and value proposition. As a community at Omdena, we remain excited to build a better world, move the human spirit forward and help organizations develop AI solutions for pressing social problems.

AI for wildfires

The Collaborators

I would not be able to close this article without thanking each of my colleagues in this challenge, whose dedication made everything possible: Iliana Vargas, Joon Sung Park, Sanyam Singh, Rohith Paul, Temitope Kekere, Ashish Gupta, Avikant Srivastava, Danielle Paes Barretto de Arruda Camara, Eric Massip, Kent Mok, Kritika Rupauliha, Leona Hammelrath, Nazgul Mamasheva, Nithiroj Tripatarasit, Rajashekhar Gugulothu, Rizki Fajar Nugroho, Robin Familara, Salil MishraSam Masikini, Tanya Dixit, Shaun Damon, Yash Bangera, Abhishek Unnam, Alexandr Laskorunsky, Amun Vedal, Angelo Manzatto, Billy Zhao, Carson Bentley, Lukasz Murawski, Sahand Azad, Yang Gao, Mikko Lähdeaho, Alyona Galyeva, Ana Maria Lopez Moreno, Brian Cerron, Gary Diana, Kennedy Kamande Wangari, Lasse Bøhling, Poonam Ligade, Serhiy Shekhovtsov, Kumar Mankala, François‑Guillaume Fernandez and Yemissi Kifouly. You can connect with me via LinkedIn.

Want to work with us too?

Let’s see if we are a good fit