AUS / NZ 2020 DATA QUEST Challenge
DATA QUEST CHALLENGE AREA : BUSHFIRES
Can AI, primed with data from multiple satellites and local sensor networks, detect fires earlier, predict fire behaviour, and help emergency services respond more effectively to protect homes, people and valuable natural capital?
As our planet warms, bushfires are an ever increasing global problem. With vast areas of forest destroyed annually, wildfires release carbon and wipe out irreplaceable ecosystems. The Sentinel-3 World Fire Atlas recorded 79,000 wildfires in August 2019, compared to just over 16,000 fires during the same period in 2018.
Predicting the behavior of bushfires is a hugely difficult problem, made more complicated by a myriad of factors such as fuel load, atmospheric conditions, soil moisture, and availability of water. Surprisingly, prediction tools are decades old and much of the task of planning on-the-ground bushfire response relies on the experience and instincts of fire-fighters - who are often volunteers.
Accurate and timely detection and mapping of wildfires is also difficult to do, especially since many fires start in remote areas. Can we fuse data from multiple sources to build more accurate maps of fires from the moment they ignite?
The four data quest teams are focused on three challenges that are connected/adjacent but different (and that use the same data sets:
Improved fuel assessment
Can we predict how long a fire will burn and what extent of the landscape it will consume? The Fuel Assessment Team, are developing an artificial intelligence based method that fuses historical fire events, ground-truth data, weather and satellite images to predict the size of a fire scar and the duration of the fire from the moment it ignites. The goal is a tool which enables preventative measures prior to ignition, target fire fighters and other response teams, or investigate the behaviour and ignition of fires and more.Early detection (and improving the identification of false positives)
Can we use high resolution models and data fusion to create predictive map using lower resolution/higher cadence, multiple data sources? There is a growing need to detect bushfires as close to ignition as possible and to accurately determine the location of the fire to improve the speed of emergency response. In remote areas, satellites may be able to detect fires long before they are seen and reported by people. The Detection Team will look at using image enhancing techniques from astronomy and machine learning to improve bushfire detection from the geostationary Himawari-8 satellite.ML fire behaviour models
Can we use historical satellite imagery to improve predictive models of the behavior of wildfires and, in turn, better inform fire risk management and response? The moisture content of vegetation is an important factor in understanding fire behaviour, both in the context of managing prescribed burns and predicting how wildfires might spread. Currently, this information is only available at low resolution, which prevents detailed insight into potential fire behaviour. The Fire Behaviour Team will apply machine learning to higher resolution satellite data to provide smaller-scale estimates of moisture content, aiding bushfire management.
What kind of challenge will suit the Data Quest?
We aim to solve difficult problems using machine learning, often combining multiple types of data and operating on the forefront of technology.
During April and May the Data Quest faculty will work together to solidify these challenges appropriate for the sprint and ensure they are matched with data sources and a clear line of sight to application or use.
What does a ‘challenge’ mean in this context?
A ML sprint’s results are grounded in isolating strong challenges at the beginning of the process, building a diverse community around research teams to accelerate progress and having strong feedback loops and partners to translate results into practice. The process moves through identifying candidate challenges through to refining those into a short list of confirmed challenges that are then verified.
A ‘challenge’ is not just something new and difficult which requires effort and determination to resolve, tackle or complete. It is a specific problem, need or desired outcome that could be tackled by interdisciplinary teams of subject or topic experts and ML experts.
The cycle starts with challenge definition. Early in the process, bringing together some of the brightest and best minds we can find, from subject experts, AI and technology, and applications to explore the challenge areas. We aim to identify some broad challenges, which the research teams could tackle. Through a process of iteration with faculty and emerging team leadership, they are refined and narrowed until each challenge has identified one, or several, tightly articulated questions to resolve.
Challenges must represent a clear and present scientific problem, for which there is available data that could be significantly advanced by AI tools and techniques. The broad challenge areas we start the cycle with move from provisional to confirmed as we understand how, and when, they meet these criteria.