DeepOE: A physics-informed neural network for atmospheric temperature profile retrieval

Existing retrieval methods based on infrared hyperspectral data still face limitations. On the one hand, physics-based iterative algorithms require multiple radiative transfer calculations, resulting in high computational costs. On the other hand, neural network approaches lack physical constraints, making it difficult to ensure physical consistency and to reasonably evaluate retrieval uncertainties. To address these limitations, a physics-informed network, DeepOE, is introduced, which integrates deep learning with the widely used Optimal Estimation Method (OEM). Synergistic observations from the ground-based Atmospheric Emitted Radiance Interferometer and space-based Geostationary Interferometric Infrared Sounder are employed to conduct atmospheric temperature profile retrieval experiments. DeepOE is capable of simultaneously providing the retrieval results and their associated uncertainties within a very short computation time. With radiosonde observations serving as the reference truth, the results show that for temperature profile retrievals in the 0–12 km range, DeepOE achieves a layer-averaged RMSE of 1.389 K, representing reductions of 0.03 K and 0.05 K compared with OEM and the purely data-driven ResNet model, respectively. 摘要 现有基于红外高光谱数据的大气参数反演方法仍存在一定局限性: 一方面, 基于物理的迭代反演算法依赖多次辐射传输计算, 计算代价较高; 另一方面, 神经网络方法缺乏物理约束, 难以保证反演结果的物理一致性及不确定度的合理评估. 本文提出了一种物理信息嵌入的深度反演框架DeepOE, 其将深度学习与最优估计方法相结合. 利用地基大气发射辐射干涉仪与星载干涉式大气垂直探测仪的协同观测, 开展了大气温度廓线反演实验.DeepOE能够在极短时间内同时给出反演结果及其不确定度估计. 以探空观测作为真实性参考, 在0–12km温度廓线反演中, DeepOE层平均均方根误差为1.389K, 较最优估计方法和纯数据驱动ResNet模型分别降低了0.03K和0.05K.