Fighting forest fires with AI in French-speaking Switzerland

Given its complexity, most simulation models of fire dynamics only work for specific locations. A team from Unit 8, specialising in developing data analysis tools with machine learning, an artificial intelligence technology based on big data, have created an algorithm based on a mix of open and proprietary data that could be applied to a wide range of situations.

Why it matters. The world has lost nearly 100 million hectares of forest in two decades, accelerating the loss of biodiversity. Forest fires are responsible for about 15% of greenhouse gas emissions.

Read more: Are 'megafires' becoming the new normal?

In 2019, an area of tropical forests the size of 30 football stadiums disappeared every minute, often due to fires set intentionally to clear land for agriculture. The Amazon alone lost the equivalent of half the surface area of Switzerland in 2019 partly due to such fires.

The project. Unit 8 started a collaboration last year with the Bolivian branch of WWF and the Fondo Ambiental Nacional del Ecuador (FAN), an NGO that funds environmental conservation projects.

According to Marcin Pietrzyk, co-founder of Unit 8:

“These organisations have ground sensors and satellite images that can detect when a fire starts. But they do not have a system that allows them to predict the spread of these fires or to detect fires that have started outside the Bolivian border.”

By 2019, such fires had led to the destruction of a million hectares in Bolivia's Chiquitania, the largest dry tropical forest in the world.

How the technology works. The reaction time of firefighters is crucial to reduce the impact of fires and to also determine which areas to protect or evacuate first. As such, Unit 8's computer scientists apply their learning machine technologies to predict both the speed and the direction of spread.

Since fires are influenced by different factors (origin of the fire, type of wood, presence of humans in its path, topography, weather, etc.), predicting its spread can be very complex.

Unit 8 is not the first to attempt predictive modelling of a forest fire. But most models are local and based on specific geographical criteria that are difficult to generalise.

There are more universal simulation models such as Farsite. However, according to Marcin Pietrzyk:

“Farsite assumes constant environmental conditions, so simulations become less accurate over time as conditions change in the real world. This leads to a build-up of modelling errors.”

The learning machine. Machine learning research applied to forest fires has so far focused mainly on detecting or estimating fire vulnerability, and less on predicting the occurrence and behaviour of forest fires.

However, Conservation International has developed FireCast, a tool for detecting and warning tropical forest fires based on satellite images, that uses convolutional neural networks to predict the spread of fires. The algorithm is designed to predict the perimeter of the fire based on the area already burned, the topographic model and meteorological variables, and to send alerts.

Unit 8 engineers have found that this research is limited by the processing of a very large amount of data from different sources. To get around this difficulty, they simplified the solution by using coarser data to reduce the computing resources required. They particularly used the Google Earth Engine geospatial data aggregation tool, which collects data in real or near real-time to detect thermal anomalies.

Unit 8 is currently enriching the data to adapt them to the case of Bolivia, by including some specific data such as soil composition.

The Unit 8 team is also experimenting with different machine learning approaches such as convolutional neural networks. Its latest approach is to use methods based on recurrent neural networks which are used for video image prediction.

The start-up. Launched in 2017 by former Palantir and Swisscom employees, Unit 8 has around 60 employees, 50 of whom are based in Switzerland.

• The company has developed around 50 data analysis projects based on machine learning for large companies such as Firmenich, Daimler as well as pharmaceutical companies, a leading insurance company and a major Swiss bank.

• As attracting talent is key in the highly competitive field of artificial intelligence, the start-up encourages its employees to work pro bono on projects of general interest.