Temperature Profile Retrieval from Ground-based Hyperspectral Observation with an Optimal Estimation driven Unrolled Neural Network

The vertical distribution of the temperature within atmospheric boundary layer plays a crucial role in meteorological research. The Atmospheric Emitted Radiance Interferometer (AERI) enables high-accuracy and high-temporal-resolution observations of downwelling atmospheric spectral radiance, from which the evolution of local atmospheric temperature profiles can be retrieved. However, existing retrieval methods suffer from limitations. On the one hand, iterative optimal estimation methods (OEM) require repeated utilizing radiative transfer model (RTM) to compute simulated radiances, leading to heavy computational costs and latency that preclude real-time applications; on the other hand, standard deep learning methods typically cannot provide posterior uncertainty estimates alongside the retrievals. For achieving real-time retrieval of atmospheric temperature profiles below 3 km from AERI hyperspectral infrared observations, we propose an OEM driven deep unrolling retrieval model that explicitly unrolls the linearized Gauss–Newton iterations in the OEM framework into a finite number of differentiable blocks with trainable parameters. Each block preserves theory guided data consistency and OEM’s inherent prior regularization while remaining capable to end-to-end learning. Crucially, the network outputs both the temperature profiles and posterior uncertainty derived from the posterior covariance, thereby combining interpretable uncertainty quantification with the computational efficiency of deep neural networks. The retrieval calculation time is reduced from the minutes typically required by the conventional OEM to under 5 seconds. Based on ERA5 reanalysis data and radiosonde observations, the retrieval performance of the proposed model was compared with that of the OEM and the Temporal Convolutional Network (TCN) model. Results indicate that the proposed model achieved the best performance in Root Mean Square Error (RMSE) and retrieval uncertainty. The proposed model exhibited its best retrieval performance in summer, achieving a layer-averaged RMSE of 0.931 K, representing reductions of 0.099 K and 0.111 K compared with OEM and TCN, respectively, and a layer-averaged uncertainty of 2.466 K, approximately 0.14 K lower than that of OEM. This research demonstrate that the proposed deep retrieval model offers an efficient and interpretable solution for real-time retrieval and uncertainty quantification of AERI hyperspectral observations, and it has the potential to further advance the application of infrared hyperspectral remote sensing in atmospheric science.