Publications

Campagnolo, ML; Libonati, R; Rodrigues, JA; Pereira, JMC (2021). A comprehensive characterization of MODIS daily burned area mapping accuracy across fire sizes in tropical savannas. REMOTE SENSING OF ENVIRONMENT, 252, 112115.

Abstract
Daily global burned area mapping of Earth from remotely sensed data is of considerable interest and indicates a clear user need, despite well-known accuracy limitations. Many studies report global and regional indicators of spatial and temporal accuracy to evaluate the quality of burned area products, but tend to rely on relatively sparse validation data. In this paper, an extensive validation data set covering 223 million ha over the Brazilian Cerrado for the 2015 fire season is used to thoroughly evaluate how the accuracy of burned area products relying on distinct spectral channels (SWIR+NIR versus NIR) and spatial resolution (250 m, 500 m, and both) depends on fire size. Toward this end, we consider publicly available burned area products MCD64A1 (v006) at 500 m and FireCCI51 at 250 m and use a recently proposed graph-based patch algorithm, which is flexible enough to combine 250-500 m bands. he trade-off between spatial resolution and spectral channels for the problem at hand reveals that better spatial resolution (NIR at 250 m) is preferable to a broader range of spectral channels (NIR + SWIR at 500 m) for small burns ( 250 ha), but that this relation is not preserved for large burns ( 1000 ha). Also, this study shows that combining both spectral indices into a single classifier leads to the most accurate burned area map. The analysis by land cover type indicates that overall accuracy is highest for grasslands and savannas, with grasslands benefiting the most from the use of SWIR bands. Results also indicate that burned areas can be classified with accuracy measured by the Dice Coefficient close to 80% for fire sizes larger than 500 ha, which corresponds to 60% of the total burned area, but accuracy lower than 36% for fire sizes smaller than 250 ha. In particular, our analysis suggests that the validation of burned area products always should be linked to the description of fire size distribution since it is a fundamental driver of accuracy. Results show that MODIS daily burned area mapping at 250-500 m spatial resolution can exhibit high temporal accuracy, with almost unbiased estimates and standard deviation close to 1.5 days, indicating an accurate characterization of the spread of large fires.

DOI:
10.1016/j.rse.2020.112115

ISSN:
0034-4257