11:45 AM - 12:00 PM
[AAS07-11] 3D Atmospheric Structure Estimation via Simulation-Observation Data Fusion For Radiative Closure Assesment
Keywords:3D Atmospheric Structure, Machine learning, Numerical simulation, Radiative closure
Achieving radiative closure, one of the primary objectives of EarthCARE, necessitates an accurate and high-resolution three-dimensional atmospheric structure. The atmospheric lidar (ATLID) and cloud profiling radar (CPR) onboard EarthCARE provide detailed vertical profiles of the atmosphere. However, their observational coverage is limited to the satellite's nadir track. Therefore, reconstructing the full three-dimensional atmospheric structure, including horizontal inhomogeneities, requires an estimation approach that integrates additional observational data. This study proposes a method for estimating the three-dimensional atmospheric structure using optical observations from EarthCARE’s Multi-Spectral Imager (MSI).
In this study, we conducted experiments to estimate the three-dimensional atmospheric structure using multi-spectral radiance data computed by a three-dimensional radiative transfer model based on an atmospheric field generated by Large Eddy Simulation (LES), which serves as a virtual ground truth. To enhance estimation accuracy, we incorporate past three-dimensional atmospheric fields evolved over time via numerical simulation as auxiliary data. A supervised machine learning model is trained using multi-spectral radiance data and numerically simulated three-dimensional atmospheric fields as inputs to accurately estimate the three-dimensional distribution of two key atmospheric radiative profiles: the extinction coefficient (EXT) and the cloud effective radius (CER). In our experimental setup, the horizontal resolution of the reference three-dimensional atmospheric field is degraded by half to substitute numerical simulation uncertainties and used as an input. The machine learning model is based on the UNet architecture, commonly employed for semantic segmentation. The simulated atmospheric fields and radiance data are encoded separately using convolutional neural networks, after which their extracted features are combined and decoded to reconstruct the atmospheric field. Estimation accuracy is evaluated using the correlation coefficient and relative error.
As a result, the proposed method achieved a correlation coefficient exceeding 0.99 for both the EXT and CER. The relative errors were 3.4 % for the EXT estimation and 5.5 % for the CER estimation, demonstrating that the three-dimensional spatial structure of clouds was accurately reconstructed. This result were showed that the correlation coefficient was higher by 0.44 for the EXT and 0.12 for the CER compered to the low-resolution data used as simulated data. The proposed method contributes to the achievement of radiative closure by enabling high-accuracy three-dimensional atmospheric field estimation based on limited observational data.
In this study, we conducted experiments to estimate the three-dimensional atmospheric structure using multi-spectral radiance data computed by a three-dimensional radiative transfer model based on an atmospheric field generated by Large Eddy Simulation (LES), which serves as a virtual ground truth. To enhance estimation accuracy, we incorporate past three-dimensional atmospheric fields evolved over time via numerical simulation as auxiliary data. A supervised machine learning model is trained using multi-spectral radiance data and numerically simulated three-dimensional atmospheric fields as inputs to accurately estimate the three-dimensional distribution of two key atmospheric radiative profiles: the extinction coefficient (EXT) and the cloud effective radius (CER). In our experimental setup, the horizontal resolution of the reference three-dimensional atmospheric field is degraded by half to substitute numerical simulation uncertainties and used as an input. The machine learning model is based on the UNet architecture, commonly employed for semantic segmentation. The simulated atmospheric fields and radiance data are encoded separately using convolutional neural networks, after which their extracted features are combined and decoded to reconstruct the atmospheric field. Estimation accuracy is evaluated using the correlation coefficient and relative error.
As a result, the proposed method achieved a correlation coefficient exceeding 0.99 for both the EXT and CER. The relative errors were 3.4 % for the EXT estimation and 5.5 % for the CER estimation, demonstrating that the three-dimensional spatial structure of clouds was accurately reconstructed. This result were showed that the correlation coefficient was higher by 0.44 for the EXT and 0.12 for the CER compered to the low-resolution data used as simulated data. The proposed method contributes to the achievement of radiative closure by enabling high-accuracy three-dimensional atmospheric field estimation based on limited observational data.