Fusion of multi-scale and multi-sensor remote sensing for fuel type classification

 Wildfires have significant socio-economic and ecological impacts, posing risks to human lives, buildings, ecosystem services and biodiversity values. Areas with increased danger are being pushed north by climate change, resulting in an additional exposure of up to 15 million Europeans. However, integrated risk assessment systems for wildfires are still lacking in many industrialized countries. This project proposal aims to facilitate bridging this gap by focusing on the accurate mapping of wildland fuel types in Belgium using diverse remote sensing techniques, exploiting the long experience obtained from similar ecosystems in Greece. Wildland fuels and their status (in terms of structure, vegetation composition, senescence and water content) are a prerequisite for any type of wildfire risk assessment framework, since they represent the biomass available for burning under the current meteorological conditions. Traditionally, fuel types have been estimated through time-consuming field observations, but recent advancements in satellite remote sensing and UAV (unoccupied aerial vehicle) technology in combination with multi-scale deep learning offer new opportunities. In particular, the design and implementation of novel methodologies for mapping fuel types based on the fusion of 3D point clouds and spectral data derived from UAVs – which is a research area that only recently has started being extensively investigated – offers great potential. 

This project specifically aims to exploit the fusion and harmonization of multi-scale remote sensing (including satellite, aerial, mobile and UAV), along with a multi-sensor (LiDAR, multispectral and RGB) approach. The hypothesis is that this fusion in combination with novel, multi-scale deep learning algorithms (including multi-scale convolutional neural networks) can support the accurate mapping of wildland fuels. While UAV- and mobile platforms can collect data at very high spatial resolution, they are limited in terms of areal coverage. The multi-scale analysis that will be performed will inform about the capacity of extrapolating region- and time-specific data to extended areas. In this context, we will consider satellite multispectral sensors (i.e., Sentinel-2), synthetic aperture radar (SAR) data (i.e., Sentinel-1) and satellite LiDAR sensors (i.e., GEDI) in terms of medium-resolution data sources. As an intermediate between satellite- and close-range remote sensing data, this project can make use of a recently collected airborne LiDAR and multispectral database covering the province of Limburg, Belgium. Three pilot study sites situated within wildland-urban-interface (WUI) areas will be identified as suitable grounds for the methodologies, located in Belgium and Greece. In these areas, field campaigns including (multispectral, LiDAR and RGB) UAV flights and mobile LiDAR missions will be carried out. 

In line with the overall objective, the project will generate detailed fuel type maps using deep learning algorithms applied to close-range (UAV and mobile) remote sensing (Work package 1) and a combination of satellite, aerial, mobile and UAV imagery (Work package 2). The accuracy of the fuel type maps will be validated (Work package 3). The results will be disseminated to the scientific community, policy makers, and land managers (Work package 4) enabling them to prioritize fuel treatments and management practices to reduce exposure to wildfires. Such information is crucial for stakeholders in order to effectively and efficiently allocate resources and implement preventive measures in high-risk areas.