Characterizing the spatio-temporal distribution, detection, and prediction of aerosol atmospheric rivers on a global scale.

Aerosol Atmospheric Rivers (AARs) are elongated and narrow regions that carry high concentrations of aerosols (tiny particles suspended in the atmosphere) across large distances, exerting effects on both air quality and human health (Chakraborty et al., 2021, 2022). Monitoring and modeling these aerosols present distinct challenges due to their dynamic nature and complex interactions within the atmosphere. In this context, the present study detects and predicts the AARs using MERRA-2 reanalysis datasets with their seasonal climatology of key aerosol species, including Black Carbon (BC), Dust (DU), Organic Carbon (OC), Sea Salt (SS), and Sulphates (SU). The study employs an innovative Integrated Aerosol Transport (IAT) based AAR algorithm from 2015 to 2022. A total count of 44,020 BC AARs, 13,280 DU AARs, 21,599 OC AARs, 17,925 SS AARs, and 31,437 SU AARs were detected globally. The seasonal climatology of BC and OC AARs intensifies in areas such as the Amazon rainforest and Congo during AMJJAS (April-September) due to forest fires. Similarly, DU AARs are more frequent in regions near the Saharan desert, primarily around the equator during AMJJAS. SS AARs tend to predominate over the oceans, while SU AARs are predominantly found in the northern hemisphere, primarily due to higher anthropogenic emissions. Furthermore, convolutional autoencoder-based models were developed for key aerosol species, strengthening predictive accuracy by effectively capturing complex data relationships and delivering precise predictions for the last 5-time frames. During validation, the model evaluation parameters for image prediction such as the Structural Similarity Index ranged from 0.86 to 0.94, Peak Signal-to-Noise Ratio fluctuated between 1.14 and 42.25 dB, Root Mean Square Error varied from 2.39 to 296.4 mg/(m-sec), and Mean Square Error fell within the range of 1.55-17.22 mg/(m-sec). These collectively reflect image closeness, quality, dissimilarity, and accuracy in AAR prediction. This study demonstrates the effectiveness of advanced machine and deep learning models in predicting AARs, offering the potential for advanced forecasting and enhancing resilience in high-aerosol concentration regions.