Construction and preliminary evaluation of a microwave radiation imager land surface assimilation observation operator based on the MLP method

In land surface brightness temperature assimilation studies, conventional radiation transfer models as observation operators often require a large amount of auxiliary information. However, the complexity and variability of land surface parameters can lead to significant errors in the auxiliary information, which severely reduces the accuracy of brightness temperature simulation and, consequently, affects the assimilation results. To further improve the direct assimilation of microwave radiation imager (MWRI) brightness temperatures from the FY-3D satellite, this study attempts to build an observation operator based on a Multi-Layer Perceptron (MLP) that does not explicitly include surface emissivity, addressing the spatial and temporal variability of land surface radiation. By introducing a land surface type grouping modeling strategy and a daytime-nighttime classification modeling method, the data dimensionality is reduced to enhance the accuracy of the assimilation observation operator. Evaluation results show that the MLP model significantly outperforms the RTTOV model in terms of simulation accuracy across most land surface types. The improvement is most notable in barren land areas, where the Mean Absolute Error (MAE) decreases from 7.297 K to 4.021 K, a 44.9% reduction, and the Root Mean Square Error (RMSE) decreases from 9.029 K to 5.721 K, a reduction of 36.6%. In addition, improvements of 38.6% and 37.3% in MAE are achieved in grassland and broadleaf forest regions, respectively. The MLP model shows the most significant accuracy improvement under daytime conditions, while it still performs better at night, although the improvement in errors is relatively smaller. The quantitative analysis results using Shapley Additive Explanations (SHAP) demonstrate that the MLP model effectively replicates the physical mechanisms of land surface radiation and holds promising practical application prospects.