The availability and temporal dynamics of vegetation biomass, or living and dead fuel, is a main driver for the occurrence, spread, intensity and emissions of fires. Several studies have shown that the fuel build-up in antecedent (wet) seasons can increase burned area in the following (dry) season. In order to estimate fire emissions, the amount and dynamics of fuel loads are commonly estimated using biogeochemical models. Alternatively, data-driven approaches to model fire dynamics using machine learning methods make often use of satellite time series of leaf area index (LAI), the fraction of absorbed photosynthetic active radiation (FAPAR), or of vegetation optical depth (VOD) as proxies of the temporal dynamics in fuel availability. Although LAI or FAPAR time series provide information about the temporal dynamics in vegetation and hence fuels, they cannot be directly used to estimate fire emissions. Alternatively, global or continental maps of fuel beds or maps of above-ground biomass such as from ESA’s Climate Change Initiative (CCI) provide direct estimates of fuel loads for different fuel types, however, they do not provide sufficient temporal coverage in order to assess temporal changes in fuel loads. Here we propose a novel data-driven approach to estimate the temporal dynamics of vegetation fuel loads by combining various Earth observation products with information from databases of ground observations.

Our approach combines the temporal information from LA, FAPAR and VOD time series and from annual land cover maps with the time-invariant information from maps of above-ground biomass and large-scale fuel data bases. Specifically, we are using the LAI and FAPAR from Sentinel-3 and Proba-V, VOD from the VODCA dataset and from SMOS, annual land cover maps from ESA CCI, maps of above-ground biomass (AGB) from ESA CCI and information from the North America Wildland Fuel Database and the Biomass and Allometry Database.

The estimation of fuel loads is based on two different approaches. The first approach makes use of an empirical allometry model to estimate the fuel loads of different biomass compartments of trees and herbaceous vegetation by using total AGB and LAI as input. Based on allometric equations the biomass of stems, branches, leaves, and total woody biomass are estimated. Thereby LAI serves a s a proxy for the temporal dynamics in leaf and herbaceous biomass. Long-term changes in total AGB estimated based on regional non-linear regressions between the spatial patterns of AGB and tree cover, maximum LAI and VOD as predictors. As alternative, the use of novel products of AGB changes such as from BIOMASCAT are explored. The allometric parameters are estimated from the Biomass And Allometry Database.

The second approach makes use of machine learning models to transfer the measurements from the North America Wildland Fuel Database to other regions. Land cover, LAI and AGB from the Earth observation datasets are used as predictors for fuel loads of trees, shrubs, grass, fine and coarse woody debris and duff. Spatial cross-validation is used to estimate, evaluate random forest regression models and to provide uncertainty estimates of fuel loads. The approaches are developed and tested in four study regions: in Brazil, southern Africa, central Asia, and northern Siberia to cover a wide range of ecosystems. First results demonstrate the feasibility to estimate temporal changes in

fuels loads by integrating the respective temporal and spatial information from various Earth observation datasets.

For this work, we acknowledge the European Space Agency for funding of the Sense4Fire (sense4fire.eu) project.