A spatio-temporal deep learning model for enhanced atmospheric correction

Atmospheric correction (AC) is essential for accurate surface reflectance (SR) prediction, as it accounts for distortions in reflectance caused by atmospheric elements such as gases, aerosols, and water vapor. AC has diverse applications, including environmental monitoring, agricultural assessment, and climate studies. Traditional physics-based models for AC are often complex, require extensive calibration, and depend on atmospheric parameters, which can be challenging to obtain. In contrast, deep learning (DL) models offer simpler, more flexible, and extensible alternatives that exclusively rely on remote sensing satellite data. However, existing DL models for AC typically consider only spatial features, overlooking temporal variations in atmospheric conditions that are ucial for enhancing SR prediction accuracy. This paper introduces the Spatio-Temporal Atmospheric Correction (STAC) model, which integrates both spatial and temporal information to improve SR predictions. STAC outperforms the state-of-the-art Season-aware Atmospheric Correction Network (SAAC-Net), achieving a 30% reduction in root mean square error (RMSE) across multiple spectral bands. The model’s generalisation ability is further demonstrated by evaluating diverse land cover types, where STAC achieved an average RMSE of 0.036, compared to 0.044 for SAAC-Net. Additionally, an assessment using the radiometric calibration network (RadCalNet) ground-measured dataset reveals that STAC achieves a mean relative difference (MRD) of 0.027, significantly outperforming both DL-based SAAC-Net (0.083) and physics-based land surface reflectance code (LaSRC) model (0.073). These results underscore the importance of incorporating temporal dynamics in DL-based AC models for enhanced atmospheric correction.