Amélioration de la modélisation des risques d'incendies de brousse pour protéger les communautés victoriennes

The Challenge

While the 2019-20 Australian bushfire disaster unfolded, DEECA (Department of Energy, Environment and Climate Action) was managing statewide bushfire risk using its 'Residual Risk' metric, calculated by the Phoenix RapidFire modelling framework. After four years in operation, the system needed a significant injection of cloud technology and machine learning capabilities to continue reducing bushfire risk across Victoria.

The Phoenix RapidFire model, originally designed for desktop use, was struggling to scale. Manual post-processing workflows created bottlenecks that prevented the system from running at the resolution required for accurate prediction. Outdated data layers limited modelling accuracy, and the existing infrastructure could not integrate new data sources such as modern weather scenarios or varied ignition patterns. High operational costs came with limited capacity. The platform needed to be accessible to a range of operators, from scientists to fire chiefs, who applied the resulting metrics in different and important ways.

DEECA needed cloud-based data storage and processing at many orders of magnitude greater than had previously existed. The volume of data required for high-resolution bushfire modelling across all of Victoria far exceeded what any on-premise system could handle. An upgrade of this scale would not be likely for several years, so the project needed to be as advanced and fit-for-purpose as possible, delivering near-limitless scale while maintaining cost-effectiveness and enabling advanced machine learning integration.

The Approach

Kablamo was engaged to design and deploy CloudFARM, a cloud-native, fully scalable AWS platform to underpin the next generation of DEECA's risk modelling. The solution was architected around four key principles:

• Cloud-Native: migrated Phoenix RapidFire to a high-performance AWS cloud environment, replacing the desktop-bound processing model with a serverless architecture
• ML Integration: Amazon SageMaker for sophisticated analysis and enhanced predictions, enabling more accurate risk assessments through machine learning
• Automation: serverless pipelines using Lambda and Athena for massive scale, eliminating the manual post-processing workflows that had previously created bottlenecks
• Cost Control: auto-scaling resources with rich operator visibility, ensuring the platform only consumed resources when running simulations

The bushfire modelling uses Phoenix RapidFire, a fire behaviour simulator that puts Australia at the forefront of bushfire tools and analysis innovation. To improve accuracy, Phoenix RapidFire was supported with cleaner data from geospatial, temporal, and historical sources. The platform allows scientists and fire chiefs to swiftly create, simulate, and repeat complex scenarios with high-resolution data. A rich user interface was introduced that allowed users to quickly create scenarios for simulation using predefined or newly uploaded data, and to repeat previous runs with slight modifications.

Advanced machine learning capabilities using Amazon SageMaker enable more sophisticated analysis and enhanced predictive accuracy. The architecture uses Amazon Athena and S3 for data pipelines, and combines Phoenix RapidFire with Bayesian Network models for prediction. Python-based processing pipelines handle data transformation and model orchestration. The entire solution prioritises scalability, automation, cloud-native design, and cost-effectiveness, ensuring that DEECA can run thousands of simulations across Victoria without manual intervention or infrastructure constraints.

The Results

The CloudFARM solution increased modelling ignition resolution by a factor of approximately 6x and simulation resolution by 2x, with the combined effect on weather scenarios delivering an approximately 70x improvement compared to the previous implementation. The fully automated, AWS cloud-native architecture has eliminated previous scaling constraints, offering near-limitless capacity in a cost-effective manner. All data pipelines are now 100% automated, removing the manual processing steps that had previously limited the system's throughput.

Before CloudFARM, running a statewide risk assessment required days of manual coordination and processing. The new platform reduces that timeline to hours, with operators able to trigger simulation runs on demand and monitor progress through the rich user interface. Scientists can compare multiple scenario outputs side by side, adjusting parameters such as ignition patterns, weather conditions, and fuel load data to understand how different variables affect predicted fire behaviour. The auto-scaling architecture means DEECA only pays for compute when simulations are running, keeping costs predictable despite the massive increase in processing capacity.

Looking Forward

This redevelopment of an important resource, the ability to predict risk and subsequently manage bushfires, is a milestone in Australia's role as a leader in bushfire management. With a suite of AWS cloud infrastructure, near-limitless data processing, and machine learning advantages, Victoria now has a powerful tool with which to save lives, property, and bushland.

The CloudFARM solution ensures that DEECA's critical risk metrics can now be produced using the most accurate and powerful digital technologies available. The platform's architecture is designed to accommodate new data sources and modelling techniques as they become available, ensuring the system remains at the leading edge of bushfire science. As new geospatial datasets, satellite imagery, and climate projections become available, they can be integrated into the modelling framework without re-engineering the underlying infrastructure. Kablamo continues to provide operational support and ongoing product development through Product Care services.