The availability of high-resolution reference datasets representing in space and time and with high accuracy areas affected by fires is strategic for the validation of remotely-sensed Burned Area (BA) products. This paper proposes a methodology designed to build a burned area reference dataset from Sentinel-2 (S2) images at continental scale by implementing a stratified random sampling scheme. Representative sample units are selected across biomes and regions with high/low fire activity; each unit covers the extent of a S2 tile (∼10 000 km2) where image time series are classified with a supervised Random Forest algorithm to extract fire perimeters by exploiting visible to near and short-wave infrared S2 wavebands at 10 to 20 m spatial resolution. Time series have to satisfy requirements on maximum cloud cover, maximum time interval between consecutive images and minimum length to be suitable for being selected and processed. The proposed methodology was applied to Sub-Saharan Africa for the year 2019 to select 50 S2 sample units where time series were processed to deliver fire reference perimeters for accuracy assessment of regional BA products. Average series length is 140 days with the longest series in the savanna biome (maximum length is 355 days, 29 consecutive S2 images) and a total of 695 S2 images were processed to build the 2019 reference dataset. This dataset was compared to burned areas derived from very-high resolution Planetscope images over five S2 tiles obtaining 15.5% omission and 11.6% commission errors. To exemplify the use of this reference dataset, S2 perimeters were used to validate the NASA MCD64A1 Collection 6 and the ESA FireCCI51 BA products. The reference dataset has been added to the Burned Area Reference Database (BARD) (Franquesa et al., 2020) and is publicly available at https://doi.org/10.21950/VKFLCH.