Refining historical burned area data from satellite observations

Sentinel-2 imagery has revealed a substantial underestimation of burned area (BA) compared with earlier sat-ellite products with coarser spatial resolution. In this context, we investigate the predictability of biases between the reference Sentinel-2 BA product developed for Sub-Saharan Africa (FireCCISFD) in 2019 and commonly used global coarse resolution BA products (MCD64, Fire CCI and C3S), providing tools to refine historical annual BA data before the Sentinel-2 era. To do so, we built a comprehensive dataset of environmental predictors of BA biases, with variables or proxies of (I) the annual BA estimated from the coarse-resolution product, (II) BA sizes, (III) the persistence and strength of BA signals, (IV) the maximum potential BA, and (V) the obstruction of land surface observation from satellites. Full and parsimonious random forest models were performed and validated through out-of-bag (OOB) estimations, and reconstructed BAs were validated with external data over space, and over time. The explained variance in BA biases was >= 78.58% (OOB) for all full and parsimonious models. The reconstructed BA data showed a high correspondence with the reference BA in the validation sites over space (>= 91.15% var. explained) and time (>= 90.37% var. explained), notably reducing biases of coarse resolution products. As an example of the model applicability, the spatial patterns of Madagascar's BA were reconstructed for 2005, 2010, 2015 and 2020, revealing a burned extent between two and four times higher than previous estimations. The proposed models are operational solutions to obtain regional and global virtually unbiased BA estimates since 2000.

Automated summary

Sentinel-2 imagery reveals a significant underestimation of burned area compared to previous satellite products. This study investigates the predictability of biases between different BA products and develops tools to refine historical BA data. Random forest models are used to reconstruct BA data, which show a high correspondence with reference data, reducing biases. The proposed models provide operational solutions for obtaining virtually unbiased BA estimates since 2000.