Optimizing GAIN model to improve AOD imputation using MODIS MAIAC data and multi-source data fusion as an example

Aerosol Optical Depth (AOD) is a crucial parameter for monitoring air quality and studying atmosphere behavior, but satellite-based AOD measurements often suffer from significant gaps due to cloud cover and other obstructions. An imputation model is needed to address large-scale Missing Not At Random (MNAR) missingness with uncertainty quantification. Generative Adversarial Imputation Networks (GAIN) are particularly suited for handling MNAR missingness but remain unexplored for AOD imputation. In this study, we optimized the GAIN model to leverage its adversarial framework to achieve full coverage, preserve spatial variability, and enhance model generalizability. The model was tested using the Moderate Resolution Imaging Spectroradiometer (MODIS) Multi-Angle Implementation of Atmospheric Correction (MAIAC) dataset across the Northeastern United States. The revised GAIN model was trained and evaluated using data from 2021 to 2022, and 2023. The imputed dataset achieved full coverage, maintaining a consistent overall trend, with the imputed values capturing the broad spatial features seen in the original data. The model was validated against the in-situ AOD measurement from National Aeronautics and Space Administration’s (NASA) Aerosol Robotic Network (AERONET). A correlation coefficient (R) of 0.90 was achieved with an alignment between the imputed data and observed AOD values. When comparing to baseline imputation models such as MissForest, k-NN, and MICE, the GAIN model achieved a Mean Squared Error (MSE) of 0.04, a Root Mean Squared Error (RMSE) of 0.17, a Mean Absolute Error (MAE) of 0.12, and a coefficient of determination (R²) of 0.84. Hyperparameter tuning of the revised GAIN model improved the R² to 0.94 and reduced overall error metrics. Using the 2023 Canadian wildfire example, the model successfully imputed AOD levels, capturing the sharp rise in aerosol concentrations where AOD levels exceeded 2.5 during June 6−7. This study provides a viable solution to missing values in satellite observation for improving AOD data coverage and enhancing the accuracy of follow-on air quality assessments and atmospheric studies.