Development and preliminary evaluation of a land surface brightness temperature assimilation operator for FY-3D Microwave Radiation Imager based on a MLP neural algorithm

In the research on direct assimilation of land surface brightness temperature data, conventional Radiative Transfer Models (RTM) used as observation operators often require extensive auxiliary information. However, the complexity and spatiotemporal variability of land surface parameters can easily lead to significant errors in auxiliary data, which severely degrade the accuracy of brightness temperature simulations and affect assimilation performance. To enhance the effect of direct assimilation of observations from the MicroWave Radiation Imager (MWRI) aboard the FY-3D satellite into land surface models, this study develops a data-driven observation operator based on the Multi-Layer Perceptron (MLP) architecture, which obviates the explicit parameterization of surface emissivity. This operator is grounded in adjoint sensitivity analysis, which identifies primary sources of simulation errors in conventional RTM. Given the complex spatiotemporal variability of land surface radiation, a modeling strategy with independent models for different surface types, day/night conditions, and seasons is adopted to reduce data dimensionality and further enhances the accuracy of the observation assimilation operator. Evaluation results demonstrate that the MLP operator significantly outperforms the Radiative Transfer for TOVS (RTTOV) model in brightness temperature simulation across most surface types. The most remarkable improvement is achieved over barren areas in summer: The Mean Absolute Error (MAE) decreases from 7.297 to 4.021 K (a reduction of 44.9%), and the Root Mean Square Error (RMSE) declines from 9.029 K to 5.721 K (a reduction of 36.6%). Additionally, MAE improvements of 38.6% and 37.3% are observed over grasslands and broadleaf forests, respectively. The MLP operator exhibits the most significant accuracy enhancement under daytime conditions, while it still maintains an advantage at night, the magnitude of error reduction is relatively smaller. The season-specific modeling approach ensures that the MLP operator achieves substantial improvements in brightness temperature simulation across all seasons, with the most notable gains in autumn—RMSE reductions of approximately 4.0 K are observed for various vegetation types. Quantitative analysis using the Shapley additive explanations (SHAP) method confirms that the MLP operator can effectively replicate the physical mechanisms of land surface radiation, highlighting its promising potential for practical applications.