We use multispectral information and machine learning (ML) to develop a daytime radiance ML regression tree (RT) model to improve the FRP retrievals when 4-µm saturation occurs. 13 VIIRS M-band data, which have corresponding METImage bands, and related geometry information are collected to be used as the ML model training data. The first model was developed using data from the large fire events in California, where major land cover types were forest (59%) and Savannas (29%). Over most regions of the world, this model can predict the 4-µm radiance under a range of saturation conditions with an average prediction error of ~19%, resulting in an average FRP prediction error of ~23%.
However, this model proved to generate larger FRP retrieval errors in most of Africa, the average prediction errors for radiance and FRP were 28% and 36%, respectively. In this study, we extended our ML method to Africa, where significantly different climate and ecological features are found from the California region. This is also reflected in the statistical populations of radiances from fire pixels. A total of 96,478 fire pixels were collected from the central African during December 2022, of which 12,852 fire pixels (~13% of the total) were above the METImage saturation level for the 4-µm band (≥344K). The major land cover type for the African saturated fire pixels is Savannas (94%). We followed the same approach as previously to develop a regional ML model. The average radiance prediction error for the African model was 3.6%, with the average FRP prediction error as 10.8%. There was a 25% average FRP prediction error decrease when using the African ML model to retrieve the African fire saturated FRPs, demonstrating the importance of developing the new African ML model. Both radiance ML models have been integrated into the NOAA Enterprise Fire (“eFire”) system for global FRP retrievals; further tuning using actual METImage measurements is planned.

