Burned Area Mapping with Sentinel-2 based on reflectance modelling and deep learning – preliminary global calibration and validation

Current global burned area products are available at coarse spatial resolutions (300-500 m), what leads to large amounts of errors, hindering an accurate understanding of fire-related processes. This study proposes a global calibration method for a sensor-independent burned area algorithm, previously used with 300 m Sentinel-3 Synergy data, and here implemented with 20 m Sentinel-2 MSI imagery. A binomial model that combines reflectance-based burned area predictions constrained by spatio-temporal densities derived from VIIRS active fires is calibrated using a reference dataset generated from Landsat imagery at a sample of 34 units across the globe. Preliminary leave-one-out cross-validation analyses show promisingly high accuracies (Dice of coefficient of 84.8%, commission error ratio of 13.2%, omission error ratio of 17.1% and relative bias of -4.5%), especially taking into account the mismatch of acquisition dates between reference and algorithm input data, what introduces apparent errors on the validation results.