Deriving burn severity from multispectral satellite data is a widely adopted approach to infer the degree of environmental change caused by fire. Burn severity maps obtained by thresholding bi-temporal indices based on pre- and post-fire Normalized Burn Ratio (NBR) can vary substantially depending on temporal constraints such as matched acquisition and optimal seasonal timing. Satisfying temporal requirements is crucial to effectively disentangle fire and non-fire induced spectral changes and can be particularly challenging when only a few cloud-free images are available. Our study focuses on 10 wildfires that occurred in mountainous areas of the Piedmont Region (Italy) during autumn 2017 following a severe and prolonged drought period. Our objectives were to: (i) generate reflectance composites using Sentinel-2 imagery that were optimised for seasonal timing by embedding spatial patterns of long-term land surface phenology (LSP); (ii) produce and validate burn severity maps based on the modelled relationship between bi-temporal indices and field data; (iii) compare burn severity maps obtained using either a pair of cloud-free Sentinel-2 images, i.e. paired images, or reflectance composites. We proposed a pixel-based compositing algorithm coupling the weighted geometric median and thematic spatial information, e.g. long-term LSP metrics derived from the MODIS Collection 6 Land Cover Dynamics Product, to rank all the clear observations available in the growing season. Composite Burn Index data and bi-temporal indices exhibited a strong nonlinear relationship (R2 > 0.85) using paired images or reflectance composites. Burn severity maps attained overall classification accuracy ranging from 76.9% to 83.7% (Kappa between 0.61 and 0.72) and the Relative differenced NBR (RdNBR) achieved the best results compared to other bi-temporal indices (differenced NBR and Relativized Burn Ratio). Improvements in overall classification accuracy offered by the calibration of bi-temporal indices with the dNBR offset were limited to burn severity maps derived from paired images. Reflectance composites provided the highest overall classification accuracy and differences with paired images were significant using uncalibrated bi-temporal indices (4.4% to 5.2%) while they decreased (2.8% to 3.2%) when we calibrated bi-temporal indices derived from paired images. The extent of the high severity category increased by ~19% in burn severity maps derived from reflectance composites (uncalibrated RdNBR) compared to those from paired images (calibrated RdNBR). The reduced contrast between healthy and burnt conditions associated with suboptimal seasonal timing caused an underestimation of burnt areas. By embedding spatial patterns of long-term LSP metrics, our approach provided consistent reflectance composites targeted at a specific phenological stage and minimising non-fire induced inter-annual changes. Being independent from the multispectral dataset employed, the proposed pixel-based compositing approach offers new opportunities for operational change detection applications in geographic areas characterised by persistent cloud cover.
Read full abstract