Fractional vegetation cover ratio estimated from radiative transfer modeling outperforms spectral indices to assess fire severity in several Mediterranean plant communities

Abstract

The obtention of wall-to-wall fire severity estimates through reliable remote sensing-based techniques that align with management needs is a critical factor in post-fire decision-making processes. In this paper, we novelty proposed a multi-date change detection framework based on the variation in fractional vegetation cover (FCOVER), with enough ecological sense and physical basis to be generalizable across different plant communities and burned landscapes with varying environmental conditions. This framework meets the definition of fire severity operationally used in the field as a biophysical indicator when fire effects on the understory and overstory layers are linked. The FCOVER was retrieved from Sentinel-2 surface reflectance scenes by inverting PROSAIL-D radiative transfer model (RTM) simulations using the random forest regression algorithm. FCOVER retrievals were validated in the field using burned and unburned control plots. We computed the FCOVERr metric as the ratio of post-fire to pre-fire FCOVER. We tested the relationship of the FCOVERr and the most common bi-temporal spectral indices in the literature, i.e. the differenced Normalized Burn Ratio (dNBR), the Relative dNBR (RdNBR) and the Relativized Burn Ratio (RBR), with the Composite Burn Index (CBI) measured in field plots for validation purposes in two case-study wildfires in the western Mediterranean Basin. We also calculated the transferability of FCOVERr and the spectral indices between different plant communities within each site, as well as between sites. The predictive errors of pre and post-fire FCOVER retrievals were found to be low (RMSE ≈ 10%) for the two study sites. Overall, the FCOVERr metric provided more accurate CBI estimations (R2 = 0.87 ± 0.04) than spectral indices (R2 = 0.71 ± 0.13). The CBI was linearly related with the FCOVERr metric for both sites, whereas the type of relationship with spectral indices was not consistent, which translated into better transferability performance of the FCOVERr metric (nRMSE = 14.27% ± 3.75%) than that of the spectral indices (nRMSE = 21.97% ± 8.09%), not only between different Mediterranean plant communities within sites, but also between the two sites. Spectral indices underestimated moderate to high fire severity to a greater extent than FCOVERr in the CBI field plots, and misclassified fire severity in several areas with patchiness fire effects identified in the field. The FCOVERr product proposed in this study may be a sound choice for the operational identification of priority areas for post-fire management.