Global Detection of Long-Term (1982–2017) Burned Area with AVHRR-LTDR Data

Abstract

This paper presents the first global burned area (BA) product derived from the land long term data record (LTDR), a long-term 0.05-degree resolution dataset generated from advanced very high resolution radiometer (AVHRR) images. Daily images were combined in monthly composites using the maximum temperature criterion to enhance the burned signal and eliminate clouds and artifacts. A synthetic BA index was created to improve the detection of the BA signal. This index included red and near infrared reflectance, surface temperature, two spectral indices, and their temporal differences. Monthly models were generated using the random forest classifier, using the twelve monthly composites of each year as the predictors. Training data were obtained from the NASA MCD64A1 collection 6 product (500 m spatial resolution) for eight years of the overlapping period (2001–2017). This included some years with low and high fire occurrence. Results were tested with the remaining eight years. Pixels classified as burned were converted to burned proportions using the MCD64A1 product. The final product (named FireCCILT10) estimated BA in 0.05-degree cells for the 1982 to 2017 period (excluding 1994, due to input data gaps). This product is the longest global BA currently available, extending almost 20 years back from the existing NASA and ESA BA products. BA estimations from the FireCCILT10 product were compared with those from the MCD64A1 product for continental regions, obtaining high correlation values (r2 > 0.9), with better agreement in tropical regions rather than boreal regions. The annual average of BA of the time series was 3.12 Mkm2. Tropical Africa had the highest proportion of burnings, accounting for 74.37% of global BA. Spatial trends were found to be similar to existing global BA products, but temporal trends showed unstable annual variations, most likely linked to the changes in the AVHRR sensor and orbital decays of the NOAA satellites.