Upscaling wildfire consumption using UAV-LiDAR and Sentinel-2 data: a Mediterranean case study

The benefits of upscaling methods applied to multi-source remotely sensed products aimed at enhancing the reliability of fire severity assessments across extensive burned landscapes have not been explored to date. In this context, light detection and ranging (LiDAR) scans procured by unmanned aerial vehicles (UAVs) may be suitable as a bridge tool for improving fire severity estimates outside areas not sampled by field inventories. This proof-of-concept study explores the spatial extrapolation of fire severity estimates at the wildfire scale by leveraging post-fire UAV-LiDAR metrics (local scale matching the extent of the UAV survey) as an intermediate step in the upscaling process to bridge field inventories (plot scale) with bi-temporal Sentinel-2 spectral indices. We considered total vegetation consumption, measured in calibration (n = 20) and validation (n = 15) field plots, as an individual fire severity indicator within a case-study wildfire in the western Mediterranean Basin. Ten-times repeated fivefold cross-validation resampling of ordinary least squares model showed that the vertical complexity index (VCI) derived from UAV-LiDAR scans was an accurate proxy for total vegetation consumption in the calibration field plots (R2 = 0.84; RMSE = 11.70), outperforming conventional Sentinel-2 spectral indices such as the Relativized Burn Ratio (RBR) (R2 = 0.61; RMSE = 18.16). The UAV-LiDAR VCI metric was used to generate a wall-to-wall product representing total vegetation consumption at the local scale, which served as an intermediate reference dataset (n = 128) in the upscaling procedure bridging field calibration plots with Sentinel-2 optical data. The use of this product significantly improved the extrapolation of total vegetation consumption estimates at the wildfire scale derived from the RBR index, as assessed in independent validation field plots (R2 = 0.76; RMSE = 15.21). The upscaling method not only outperformed the traditional wildfire-scale extrapolation of RBR estimates calibrated using a limited number of field inventories without the UAV-LiDAR bridge (R2 = 0.53; RMSE = 23.31), but also minimized the underestimation of high fire severity. The upscaling method using UAV-LiDAR data offers a practical approach to reducing field sampling effort while enhancing the generalizability of fire severity estimates. Future research should evaluate the applicability of this method across multiple wildfire events in Mediterranean regions and in other biomes.