Enhancing Aerosol Vertical Distribution Retrieval With Combined LSTM and Transformer Model From OCO-2 O2 A-Band Observations

Abstract

The precise determination of aerosol vertical distribution is crucial for accurate radiative transfer simulations in atmospheric aerosol studies. This research utilizes Orbiting Carbon Observatory-2 oxygen A-band hyperspectral observation data, which are sensitive to aerosol vertical distribution. We propose a novel machine learning model that combines long short-term memory and Transformer architectures. Furthermore, a physics-based, information-driven band selection method was developed to simplify input data and reduce complexity. To enhance the algorithm's applicability, the model was applied across the entire African continent and adjacent water bodies. For multiple dust events in West Africa, the retrieved aerosol layer height (ALH) and aerosol optical depth (AOD) values exhibit strong agreement with the cloud–aerosol Lidar with orthogonal polarization, yielding the correlation coefficients of 0.6893 for AOD and 0.7866 for ALH. The model's high retrieval accuracy is validated using two metrics: Earth mover's distance and mean-squared error. By integrating advanced machine learning techniques into remote sensing, this study achieves a significant improvement in retrieval accuracy over previous methods. Through systematic optimization, the model provides a robust solution for accurately characterizing aerosol layers, making it a valuable tool for advancing atmospheric aerosol research.