Spatial-Temporal Pre-trained Transformer Model (SPT) for coarse and fine mode aerosol size retrieval

The SPT model aims to examine each pixel in the satellite dataset to learn its spatial and temporal characteristics, deeply exploring the spatiotemporal relationships inherent in multi-band satellite data. Moreover, the pre-trained learning framework in SPT does not require ground station data, allowing it to use pixel data from all satellite images to learn pixel-scale spatiotemporal relationships. On the foundation of pre-training, the model then further aligns with ground-based (AERONET) data to establish robust inversion models for fAOD and cAOD. This approach maximizes the use of remote sensing data addressing a critical gap in current methodologies where large portions of satellite data remain underused due to the absence of corresponding ground measurements.