Aiming at the complex distribution characteristics of annular temperature fields, including radial gradient heterogeneity, circumferential periodic fluctuation, and local temperature inversion, as well as the deficiencies of traditional reconstruction methods in annular structure adaptability and physical consistency, this paper proposes an annular temperature field reconstruction method combining adaptive Radial Basis Function (RBF) interpolation, a Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM) hybrid network, and hierarchical physical constraints. Firstly, a hybrid-kernel RBF interpolation with radially adaptive kernel parameters is adopted to construct the initial temperature field through sparse measurement data. Then, the CNN-LSTM hybrid network is utilized to extract spatial local features and circumferential periodic features to correct the residual error of the initial temperature field. Finally, according to the heat conduction law, differentiated hierarchical physical constraints are applied to different regions of the temperature field to improve the physical rationality and credibility of the reconstruction results. Multi-condition experimental results show that the reconstruction errors of the proposed method under typical working conditions of 500 K, 1 000 K, and 1 750 K all meet the engineering error threshold of 5%. The spatial resolution reaches 0.5 mm, which is better than the engineering index of 2 mm. Ablation experiments verify that the LSTM module ensures the circumferential continuity of the temperature field, and the hierarchical physical constraints significantly enhance the physical credibility of the reconstruction results. The proposed method can provide a reliable technical support for high-precision annular temperature field reconstruction, condition monitoring, and performance optimization of aero-engines.