Time series retrieval of Multi-wavelength Aerosol optical depth by adapting Transformer (TMAT) using Himawari-8 AHI data

Accurate aerosol optical depth (AOD) with high temporal resolution is required for dynamic monitoring changes
of aerosol properties. The Himawari-8 Advanced Himawari imager (AHI) data with 10-min temporal resolution
has been widely used for AOD retrieval but usually involves radiative transfer models and assumptions on surface
reflection and aerosol properties that are hard to satisfy. This paper introduces a data-driven algorithm without
any priori assumptions on surface or aerosol properties called Time series retrieval of Multi-wavelength AOD by
adapting Transformer (TMAT). TMAT directly retrieves AOD at eight-wavelengths every 10-min throughout the
day, using AHI top of atmosphere (TOA) reflectances in 6 wavebands and ancillary data. TMAT explores diurnal
temporal change patterns of AOD and surface reflection using time series data. TMAT was trained and evaluated
using samples derived by matching a year (2019) of AHI full-disk data and AERONET/SONET observations with
12,768 diurnal time series, each of which includes up to 81 10-min 6-band AHI TOA observations and 81 (time)
× 8 (wavelength) AOD values per day. There are up to 273,413 (<12,768 × 81) TOA observations in 6 wave-
bands and up to 1,420,716 (<12,768 × 81 × 8) AOD values in all the time series samples, with filled values
( 9999) to represent no data due to cloud contamination and incomplete wavelength configurations of AOD
stations. The masking mechanism in Transformer is leveraged to exclude the contribution of the AHI filled values
in time series AOD retrieval and exclude the contribution of AOD filled values in training. TMAT was trained with
a physical constraint that the retrieved AOD decreases toward longer wavelengths. TMAT AOD was evaluated by
separating training and validation samples using site-specific leave-one-station-out validation. TMAT AOD shows
extremely high accuracy at all wavelengths, with RMSE = 0.101, R = 0.949 and 81.69% of retrievals within ±
(0.05 + 15%*AODsites) for the AOD at 550 nm. We also validated the accuracy of the NASA multiangle imple-
mentation of atmospheric correction (MAIAC) AHI AOD over these sites with RMSE = 0.186, R = 0.832 and
52.10% of retrievals within ±(0.05 + 15%*AODsites). The Ångstr¨om exponent (AE) derived from the multi-
wavelength TMAT AOD also shows good agreement with AERONET/SONET AE, with a correlation coefficient
of 0.620 and an RMSE of 0.291. The TMAT effectiveness to capture diurnal temporal change patterns of surface
reflection and AOD was evidenced by the significant reduction of AOD accuracy when using a non-time series
deep learning method. The effectiveness of multi-wavelength joint retrieval was evidenced by the significant
reduction of AE accuracy when using single wavelength retrieval models. Evidence shows the training and
evaluation split methods have a large impact on evaluation accuracy and this study further used ground network 
independent training and evaluation split to confirm the leave-one-station-out validation accuracy. Despite this,
future study will focus on more spatially independent validation and more comprehensive comparison between
the MAIAC and TMAT results.