Utilizing deep learning techniques in operational system of geo-KOMPSAT-2A satellite for estimating solar radiation and aerosol optical depth

Deep learning (DL) models have emerged as promising tools to enhance satellite observations, yet their real-time deployment in operational systems for estimating solar radiation and aerosol optical depth (AOD) remains infrequent. This study presents the successful integration and three-year operational validation of DL models within the Geo-KOMPSAT-2A (GK-2A) satellite’s high-throughput processing pipeline. We selected the back-propagation neural network (BPNN), the one-dimensional convolutional network (Conv1D) and Feature Tokenizer Transformer (FT-Transformer) as candidate models due to their reliability, stability, and runtime efficiency. Ground-validation results demonstrated the performance differences between the models. For solar radiation, the BPNN model achieved a root mean square error (RMSE) of 72.45 W/m², a normalized RMSE (nRMSE) of 17.28%, a mean bias error (MBE) of 48.34 W/m², and a coefficient of determination (R2) of 0.92. In the case of FT-Transformer model, it recorded an RMSE of 72.02 W/m², nRMSE of 17.52%, MBE of 47.28 W/m², and R2 of 0.92. Regarding AOD estimation, the BPNN model showed an RMSE of 0.053, nRMSE of 19.78%, MBE of −0.034, and R2 of 0.95. For the FT-Transformer, it exhibited improved performance with an RMSE of 0.03, nRMSE of 14.07%, MBE of 0.01, and R2 of 0.97. FT-Transformer achieved the best accuracy across the evaluation metrics, and the cross-validation results indicate that its performance remains consistent across withheld stations, suggesting robust spatial generalization and reduced sensitivity to location-specific overfitting. Accordingly, we adopt the FT-Transformer as the primary operational model for GK-2A product generation. Notably, the operational DL system processed data significantly faster—up to 12-fold for solar radiation and 43-fold for AOD compared with conventional physical models—enabling real-time performance with enhanced accuracy. Furthermore, monthly retraining and monitoring ensured sustained robustness over the study period. These findings underline that, beyond improving predictive accuracy (notably for AOD), the core innovation lies in achieving a stable, scalable, and efficient integration of DL within operational satellite systems.