Mitigating underestimation of fire emissions from the Advanced Himawari Imager: A machine learning and multi-satellite ensemble approach

Abstract

The accurate estimation of biomass burning emissions has played a crucial role in air quality and climate forecast modeling. Satellite-based fire radiative power (FRP) has proven effective for calculating biomass burning emissions. However, FRP-based emission estimations in East Asia often rely on polar-orbiting satellites owing to the unstable performance of Japan Aerospace Exploration Agency Advanced Himawari Imager (JAXA AHI) from poor detection capability and unproper FRP retrieval method. To address this, we improve the FRP by machine learning based on mid-infrared (MIR) radiance method, leveraging the superior fire detection model developed in our previous study. In addition, we propose a multi-satellite distance-based weighted ensemble FRP estimation method. Compared to traditional MIR radiance methods, the machine learning-based FRP estimation model exhibited promising performance (correlation coefficient: 1, mean bias error: 0.2, mean absolute percentage error: 1.9%). The integration of machine learning-based FRP estimation and fire detection model dramatically mitigated the underestimation issues from the JAXA AHI. The machine learning-based FRP was combined with the Moderate Resolution Imaging Spectroradiometer FRP to create a multi-satellite ensemble FRP. Comparative assessments using the TROPOspheric Monitoring Instrument and conventional bottom-up method demonstrated that the proposed method produced reliable output. Furthermore, impact analysis revealed that missing peaks or underestimated burn scars could lead to fatally low emissions; however, the proposed method was relatively robust against missing data owing to its multi-satellite ensemble. By identifying potential FRP problems and their impact on emission estimations, this study provides valuable insights for FRP-based emission estimation studies.