Data Release for the Validation of the USGS Landsat Burned Area Essential Climate Variable (BAECV) across the conterminous U.S.

The Landsat Burned Area Product Validation dataset was collected to determine the accuracy of The Landsat Burned Area Essential Climate Variable (BAECV) product, developed by the U.S. Geological Survey (USGS). The BAECV maps burned areas across the conterminous United States (CONUS) for the entire Landsat archive (1984 – 2015). Rigorous validation of such products is critical for their proper usage and interpretation. The sampling design used to derive this validation dataset was adapted from the methods used by European Space Agency’s (ESA) Climate Change Initiative (CCI) fire_cci project to generate the first statistically rigorous global reference dataset for a burned area product that meets the CEOS LPVS stage 3 validation requirements. Our validation dataset consists of 28 Landsat path/rows across the CONUS which were selected using a stratified sampling scheme across the major Olson biomes, as summarized by the fire_cci project (Olson et al., 2001; Padilla et al. 2014). Within the CONUS this included temperate forest, Mediterranean forest, temperate grassland and savannah, tropical and subtropical grasslands and savannah, and other which included desert/xeric shrub and flooded grasslands (Padilla et al. 2014). Path/rows selected within each biome were meant to represent high and low burned areas as specified by the Global Fire Emissions Database (GFED) version 3 (Giglio et al, 2009, 2010). We used systematic sampling to select 5 validation years spaced out in 5 year increments (2008, 2003, 1998, 1993 and 1988). The validation dataset was then independently generated by three different analysts. Each analyst mapped “new” burned areas using Landsat pre-fire and post-fire image pairs. The burned area polygons were generated using the Burned Area Mapping Software (BAMS), which is a semi-automated algorithm developed by the University of Alcala, Madrid, and implemented by the fire_cci project (Bastarrika et al., 2014; Padilla et al., 2014). The outputs were manually edited using visual interpretation. From these outputs, three renditions of the validation datasets were generated in which burned area extent ranged from liberal (or inclusive) (Level 1) to conservative (Level 3). Burned area extent was defined as (1) at least one analyst identified a given pixel as burned (Level 1), (2) at least two of the three analysts were required to agree a given pixel was burned (Level 2), (3) all three analysts were required to agree a pixel was burned (Level 3). Full details of the methods used to derive this validation dataset are provided in Vanderhoof et al. (2017).