Advanced Time series Multi-wavelength AOD Transformer (TMAT) model: joint retrieval of aerosol optical and microphysical properties using Himawari-8 AHI data

• Proposes an advanced Time series Multi-wavelength AOD Transformer (TMAT) for fAOD, cAOD & SSA retrieval. • Accuracy of fAOD550, cAOD550, & SSA440: RMSE = 0.095/0.039/0.042 & R = 0.955/0.762/0.552. • Spatially independent validation shows more realistic accuracy than other validations. • TMAT attention weights were found to correlate well with atmospheric path length and scatter angle. • 2019 and 2020 Himawari-8 full-disk data collocated with AERONET and SONET observation. Current approaches to retrieving aerosol properties often rely on static or single-time observations, neglecting valuable temporal dynamics, and struggle to estimate multiple parameters beyond aerosol optical depth (AOD). This study presents an advanced time series multi-wavelength aerosol transformer (TMAT) model. TMAT advances beyond AOD retrieval by jointly estimating key, yet challenging, microphysical properties—fine/coarse mode AOD (fAOD/cAOD) and single scattering albedo (SSA)—from Himawari-8 observations at 10-minute temporal and 0.02° spatial resolution. TMAT leverages the end-to-end learning capacity of the Transformer architecture to exploit temporal information in geostationary satellite data and incorporates physically consistent activation and loss functions. TMAT was trained and evaluated using two years of full-disk Himawari-8 AHI observations matched with AERONET/SONET measurements. Site-based 5-fold cross-validation demonstrated TMAT’s robust performance. For AOD at 550 nm, fAOD and cAOD, the correlation coefficients (R) were 0.952, 0.955, and 0.762; the root mean square errors (RMSEs) were 0.092, 0.095 and 0.039; and 60.02%, 61.54% and 81.80% of the retrievals met Global Climate Observing System requirements. SSA and absorption AOD retrievals at 440 nm exhibited RMSE (R) values of 0.042 (0.552) and 0.037 (0.736), respectively. For AOD (550 nm) ≥ 0.3, the R values for the derived Ångström Exponent and fine mode fraction were 0.786 and 0.817, with corresponding RMSEs of 0.199 and 0.093. TMAT’s Transformer attention weights exhibited physically consistent behavior by assigning greater weights to observations with longer atmospheric path lengths and smaller scattering angles, where such geometric conditions enhanced aerosol signal sensitivity. These results pave the way for generating multiple aerosol product using Himawari-8/9 satellite data under the TMAT framework.