Using EO and AI Algorithms to Support Wildfire Management

News
Event date: 
18 February 2021

Introduction

Wildfire is a natural component of the Earth system, important for nutrient release and vegetation growth. However, it is clear climate change is contributing to more frequent, destructive and less predictable wildfires worldwide. Through the recent extensive fire events in the USA, Australia and the Amazon rainforest, the true cost of wildfires have been seen on both a human and environmental level. This is therefore leading to increased demand for better monitoring and analysis of burned area to improve land management and mitigate the impacts of fires, as well as to better understand the drivers of fire activity and its relevance in biogeochemical cycles, climate, and air quality. In response to this demand, CGI are leading an ESA-sponsored project to develop and demonstrate a pre-operational wildfire burned area mapping service through the implementation of machine learning algorithms and utilization of EO data. The deployment of such a service trained on increasingly frequent and high-resolution satellite observations will enable wildfire management services access to higher quality, more rapid burned area mapping products.

Copernicus Sentinel-2 satellite image of fire in Australia
Sentinel-2 image during the 2019/2020 Australian fire season. Copernicus Data/ESA/Sentinel-2

With the overarching goal of the project to provide a service that helps meet the performance and operational needs of the fire monitoring community, the project aims to provide two new aspects of functionality to improve on existing burned area mapping services. Firstly, enabling improved data availability through the effective use of current operational platforms, in particular Copernicus’s Sentinel-2 data. Secondly, the use of AI to ingest this data and improve on current methods distinguishing between burned and non-burned areas. The combination of AI and extensive EO data sets will allow for fire management teams to gain a better understanding of burned area extent in their required region of interest.

How can AI help in wildfire mapping?

The application of AI to this use-case is an important potential step to enhance current solutions, as whilst considerable effort has been put into improving global burned area algorithms, accuracy levels are still only around 60% . This is because many fires are not visible to optical satellites due to cloud cover, low severity and the size of the burned area in cases where fires are small but numerous. It is therefore necessary to generate products of a higher spatial resolution to capture these details. The application of higher resolution Sentinel-2 data compared to existing MODIS / VIIRS based products will allow smaller fires to be assessed. However, to manage the variable nature of fire occurrences in different locations across the world, multiple stand-alone satellite-based algorithms are required to accurately define burned area for the range of possible scenarios. With the presence of wildfire in most countries across the world, this is an impractical and expensive approach. Alternatively, machine learning offers a more generalised approach which is characterised by the data used to train the algorithm. With a training set representing varied types and severity of fire, the data can adequately characterise the breadth of fire activity across the globe, and provide an accurate probabilistic estimate of a given pixel being burned. Furthermore, an AI-enabled service provides a much more sophisticated approach to burned area mapping by identifying patterns of spatial and spectral characteristics that indicate burned area, in comparison to previous services which simply apply a threshold on the spectral differences between two images, which in reality would not be a set value across the region of interest. This should therefore result in a much more reliable characterisation of burned area. This is however dependent on the training data, and hence the selected algorithm will be trained using a vast reference data set generated by the University of Leicester as part of ESA’s Climate Change Initiative (CCI), which provides unprecedented spatial and temporal distribution that characterise the nature of vegetation fires globally.

How to access the service

The AI-enabled burned area mapping service will be made available to the environmental community through the EO4SD Lab as an on-demand service, allowing access to the appropriate EO data and processing capabilities. As with other services in the Lab, users will be able to define input data, regions of interest, and time period to execute the service and visualise the results relevant to them. Alongside this the EO4SD Lab has a range of services applicable to fire monitoring. Most notable of which is the more traditional Burned Area mapping service created through a differenced Normalised Burn Ratio (dNBR) of two Sentinel-2 images, to give indication of burned area and severity. Other relevant services include land use maps for more effective area management and NDVI products to help detect time to vegetation recovery or, indirectly, fuel moisture conditions.

Burned area map of the wildfire near Perth, Australia from February 2021, generated using dNBR between two Sentinel-2 images and applying the USGS colour scale for burn severity
Burned area map of the wildfire near Perth, Australia from February 2021, generated using dNBR between two Sentinel-2 images and applying the USGS colour scale for burn severity. Click here to access this map.

The 8-month long project is currently in its initial phases, engaging with expected users to refine the requirements of the community and define the service specification to appropriately meet these needs. Development and test of the AI algorithm and its deployment will take place between February and May, with the service expected to be made available to users on the EO4SD Lab from May 2021 onwards. For more details please view the AI4EO Wildfire Project Technical Note.

EO WikiThe EO Wiki contains some examples of these generated products:

  • Maps of the Perth, Western Australia fire from February 2021 are available here.

Like to learn more

We encourage prospective users of the service to get in contact to find out more about the service and its development. Further to this, with the service specification still in consideration, feedback and contributions from end-users is critical in defining the nature and performance of an idealised burned area mapping service. We therefore welcome your input on whether a service such as this would be a useful addition to your tool set and how it can best serve your needs, along with regions of interest you think the service could be used.

If you would like to know more or get involved please contact us via the support@eo4sd-lab.net