Microsoft, Azure & AI Tech Combat California Wildfire Risks

According to NASA, everyone on Earth is touched by the effects of climate change, such as hotter temperatures, shifts in rain patterns and sea level rise. 

NASA says that there is unequivocal evidence that Earth is warming at an unprecedented rate and human activity is the principal cause.

As temperatures increase globally, wildfires are becoming a growing threat. 

In the US, as temperatures rise, prolonged drought and expanding development increase the risk to communities and ecosystems.

Nearly 78,000 wildfires were reported across the country last year, representing a 20% increase on the previous year. 

In California, where fire has long played a role in shaping forests, grasslands and coastal landscapes – the challenge is increasingly about managing how communities live alongside the impacts of climate change. 

Scientists, firefighters and Microsoft’s AI for Good Lab are now using AI to detect wildfires earlier and give emergency teams more time to respond.

We want this technology to be available to people across the world

Juan Lavista Ferres, Vice President and Chief Data Scientist at Microsoft’s AI for Good Lab.

Microsoft has committed US$5m to advance AI-powered wildfire detection capabilities, with the funding directed towards expanding real-time computer vision systems that could demonstrate new approaches to deploying ML in time-critical environmental monitoring scenarios.

Computer vision at scale

For decades, wildfire detection often relied on people spotting smoke from roads, lookout points or nearby communities before alerting emergency services.

ALERTCalifornia, founded at the University of California San Diego, is changing that approach through a network of nearly 1,300 cameras positioned across fire-prone areas of the state.

The system continuously monitors live camera feeds and uses AI to identify potential signs of smoke.

It can compare different camera angles, verify a fire’s location and send alerts to emergency teams within minutes.

In its first two months of operation, the network detected 77 incidents before they had been reported through other channels.

The system now processes more than seven million images every day and can identify a wildfire as early as 2.5 hours before the first 911 call.

The AI model must distinguish genuine fire signatures from environmental noise including clouds, fog and variable lighting conditions across different times of day and seasonal changes.

This classification challenge becomes more complex when considering the diversity of landscapes monitored, from dense forests to grasslands and coastal regions.

Dr. Neal Driscoll, Geophysicist and Founder, ALERTCalifornia, says: “The data isn’t just telling us what the landscape looks like. 

“It helps us understand how it changes over time.” 

According to the project’s operational data, achieving detection 2.5 hours before traditional reporting methods indicates the model’s sensitivity threshold has been calibrated to identify smoke at early stages while maintaining acceptable precision rates.

Model performance in production

The system’s deployment demonstrates practical considerations for AI systems operating in critical detection scenarios.

Processing seven million images daily requires infrastructure capable of handling continuous inference at scale, with the Azure grant component of Microsoft’s US$5m contribution potentially addressing computational requirements.

Wildfire facts from Earth.Org

• Wildfires need three key conditions to ignite: Dry vegetation and other fuel, oxygen-rich air and a source of heat or ignition can combine to start a wildfire, while strong winds can accelerate its spread.

• Lightning can trigger wildfires naturally: Lightning is a major natural ignition source, with hotter, longer-lasting strikes more likely to start fires. A 2014 study found that a 1°C rise in temperature could increase lightning frequency by 12%.

• More than 80% of US wildfires are linked to people: Human activities such as unattended campfires, cigarettes, barbecues and pyrotechnics account for about 84% of wildfires in the US.

• Wildfires can cause widespread air pollution: Smoke contains gases and fine particles that can travel long distances and penetrate deep into the lungs, increasing risks to respiratory and cardiovascular health.

• Climate change is lengthening wildfire seasons: Rising temperatures, drought and reduced rainfall are leaving more vegetation dry and flammable. In the US, fire seasons that once lasted around four months can now extend for six to eight months or more.

• Wildfires create a climate feedback loop: Large-scale fires release substantial greenhouse gas emissions while destroying vegetation that stores carbon. This can contribute to further warming, creating conditions that increase the likelihood and intensity of future wildfires.

In 2023, the system’s detection capabilities contributed to containing Kern County, US’s Trotter Fire at 52 acres, with projections suggesting the fire could have reached nearly 4,000 acres without AI-powered intervention.

These real-world outcomes offer case studies for evaluating model impact beyond standard accuracy metrics.

The technology’s role during Sonoma County, US’s Kincade Fire in 2019, supporting the evacuation of more than 180,000 residents without loss of life, demonstrates how early warning systems can influence emergency response protocols.

“We gain so much knowledge in the first five minutes,” says Zachary Wells, Deputy Chief, Kern County Fire Department and Deputy Director of Operations, ALERTCalifornia.

“That early understanding informs everything that follows.”

Deployment scalability and adaptation

Microsoft’s funding structure, comprising US$2m for technology development and a US$3m Azure grant, could indicate priorities around both model enhancement and computational infrastructure.

“We want this technology to be available to people across the world,” says Juan Lavista Ferres, Vice President and Chief Data Scientist at Microsoft’s AI for Good Lab.

Scaling the approach beyond California presents technical challenges around model generalisation and transfer learning.

Wildfire risks vary significantly between locations, meaning systems need to account for local landscapes, vegetation, weather conditions and emergency response structures.

A model trained primarily on California data may require fine-tuning or retraining when deployed in regions with different vegetation types, topography or climate patterns.

The project is also developing capabilities beyond initial detection, incorporating nearly 98,000 square miles of LiDAR data to create detailed digital landscape models.

This expansion into terrain mapping, forest health monitoring and post-fire risk assessment suggests a broader data pipeline architecture.

Combining computer vision for smoke detection with LiDAR-derived environmental data could enable multi-task learning approaches that improve overall system intelligence and provide additional use cases for the underlying infrastructure and datasets.

Fast Facts on the Climate from the United Nations

• Global surface temperature has increased faster since 1970 than in any other 50-year period over at least the last 2000 years.

• The Earth is now about 1.42°C warmer than it was in the pre-industrial era (1850-1900).

• 2024 was the warmest year on record, with the global average near-surface temperature 1.55°C above the pre-industrial baseline.

• 2015-2024 was the warmest recorded decade.

• Every fraction of a degree of warming matters. With every additional increment of global warming, changes in extremes and risks become larger.

• Carbon dioxide (CO2) is accumulating in the atmosphere faster than any time experienced during human existence, rising by more than 10% in just two decades.

Microsoft AI for Good Lab’s projects

Aurora forecasting: Microsoft’s Aurora Forecasting project uses a 1.3 billion-parameter AI foundation model to analyse atmospheric data and improve weather and climate forecasting. The model is designed to support applications including weather prediction, air quality modelling and extreme weather analysis. Its development demonstrates how AI can process complex environmental datasets to strengthen understanding of atmospheric conditions and support resilience planning.

Geospatial machine learning: Microsoft’s AI for Good Lab combines geospatial data, satellite and aerial imagery with machine learning in collaboration with universities, conservation agencies, NGOs and Turkey’s Ministry of Interior Disaster and Emergency Management Presidency. Projects include earthquake building damage assessment, glacier and land-cover mapping, poultry barn mapping and renewable energy monitoring, supporting disaster response, conservation, humanitarian action and environmental planning.

Glacier mapping: Microsoft’s AI for Good Lab uses machine learning and satellite imagery to support glacier mapping and ecological monitoring in the Hindu Kush Himalaya region. The project identifies and outlines both clean ice and debris-covered glaciers, while a web tool allows experts to review and correct model predictions. The approach aims to accelerate mapping and improve understanding of glacier ecosystems affected by climate change.

Renewable energy mapping: Microsoft’s Renewable Energy Mapping project uses geospatial machine learning to map and monitor renewable energy development at scale. The technology can help organisations understand where renewable infrastructure is being developed and track changes across large areas. By combining AI with geospatial information, the project demonstrates how machine learning can support renewable energy planning and provide data for monitoring the transition to cleaner energy.

Bioacoustics: Microsoft’s AI for Good Lab collaborates with conservation organisations and research labs to apply machine learning and deep learning to large volumes of wildlife audio. Projects include Project Guacamaya, in partnership with the Humboldt Institute, using bioacoustics for species identification in the Amazon. Other work supports beluga whale monitoring and automated classification of bird and amphibian calls, helping advance biodiversity research and conservation.

Accelerating biodiversity surveys: Microsoft’s AI for Good Lab, Microsoft Research and Microsoft AI for Earth collaborate with NOAA Fisheries, Sieve Analytics and LILA BC to accelerate biodiversity surveys using machine learning. The projects analyse imagery and audio from camera traps, aerial cameras and microphones, reducing manual annotation. This work aims to provide conservationists with wildlife population data faster, supporting decisions on habitat protection, infrastructure and anti-poaching efforts.

Land cover mapping: Microsoft’s Land Cover Mapping work uses computer vision to turn remote sensing data into land-use and land-cover information. The approach can reduce the time environmental scientists and geospatial analysts spend manually mapping areas, allowing more focus on analysis and decision-making. By automating parts of the mapping process, Microsoft is supporting environmental monitoring and helping organisations work with large-scale geospatial datasets.