Improved burn severity estimation by using Land Surface Phenology metrics and red edge information estimated from Landsat


Global wildfire activities are expected to increase substantially in the near future. Existing techniques for spaceborne burn severity estimation often rely on bi-temporal spectral indices, which are related to in-situ burn severity data. However, due to cloud coverage and limited revisit frequency, in combination with the date of field surveys, it is a challenge to find suitable and phenologically comparable pre- and -post-fire images. To overcome these issues and to improve the accuracy of burn severity estimations by incorporating ecologically relevant spectral information, we investigated the capability of using Land Surface Phenology (LSP) metrics and incorporating red edge spectral information. We examined the well-researched Jasper fire (September 2000, Black Hills, USA) with a dense time series of Landsat-5 and -7 data. We generated synthesized red edge spectral bands through a recently proposed spectral harmonization technique and computed several bi-temporal vegetation indices. Additionally, we derived various bi-annual LSP metrics from the same indices. We used linear regression between composite burn index (CBI) ground truth data and the various indices to measure the performance of each approach, and intercompared estimated burn severity maps. We found added value of both incorporating red edge spectral information into bi-temporal indices and into LSP metrics. Among the indices, NDVI and NDVIre1n performed best, with the latter being the overall winner. This was observed for both the bi-temporal indices and the bi-annual LSP metrics, wherein best estimation performance was found with Value of Peak of Season and Value of Green Mean metrics. Although the correlation between CBI point measurements and bi-temporal index data is similar to the LSP approach, the LSP-based burn severity maps show more robustness with regard to clouds and cloud shadows, altitude gradients and pre-processing uncertainty. The results are not only relevant for sensors with native red edge bands like Sentinel-2 but also suggest that back-casting the red edge spectral information to the Landsat archive combined with an LSP based estimation approach may improve existing burn severity maps, especially in more frequently clouded regions.

International Journal of Applied Earth Observation and Geoinformation