Retrieval of Atmospheric Aerosol Optical Depth From AVHRR Over Land With Global Coverage Using Machine Learning Method
Xiaoqing Tian, Ling Gao, Jun Li, Lin Chen, Jingjing Ren, Chengcai Li
Abstract
Aerosols play an important role in global climate change, which requires long-term data records. Advanced very high-resolution radiometer (AVHRR) provides continuous observations for up to 40 years since 1979, which makes it worthwhile to retrieve aerosol optical depth (AOD) from AVHRR over land. A novel algorithm for retrieving AOD from AVHRR is developed based on the machine learning (ML) method. The AVHRR observations from pathfinder atmospheres–extended (PATMOS-x) Level-2 dataset and corresponding AOD products ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$0.55~\mu \mathrm {m}$ </tex-math></inline-formula> ) from moderate resolution imaging spectroradiometer (MODIS) in 2014 are used as training data. And AOD products in three years (2015, 2006, and 1998) named AVHRR XGB-AOD were generated for evaluation. Comparisons show that the AVHRR XGB-AOD is consistent with the MODIS AOD with correlation coefficients greater than 0.80 and RMSE less than 0.18 for most months in 2015 and 2006. The temporal and spatial characteristics from AVHRR XGB-AOD are similar to those from the MODIS AOD, but those from the previous AVHRR AOD with deep blue (DB) algorithm are significantly different. Validation with AERONET indicates that more than 68% of the matchups fall within expected error [EE, ±( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$0.05\,\,\pm \,\,0.25\times {\mathrm {AOD}}_{\mathrm {AERONET}}\mathrm {)] }$ </tex-math></inline-formula> in 2015 and 2006, while the fraction is 66% in 1998. Compared to the DB algorithm, the ML-based algorithm performs better in high-AOD conditions over vegetated regions, such as in Southeast Asia, where the DB algorithm significantly underestimates. In low-AOD conditions, the ML-based algorithm performs better over western North America and Australia, where the aerosol composition varies greatly.