Effects of Deep Learning Observation Operators in Direct Radiance Assimilation of Microwave Radiation Imager in Land Surface Models

Soil moisture is a key forecast variable of land surface models. Direct assimilation of microwave brightness temperature data to optimize soil moisture initial fields is an effective approach to improve the simulation accuracy of soil moisture. However, most existing direct assimilation methods adopt physical radiative transfer models as observation operators, and their complex parametric errors greatly restrict the improvement in assimilation performance. This study introduces a high-precision MLP (Multilayer Perceptron)-based surrogate radiative transfer model as the observation operator. Combined with the Simplified Extended Kalman Filter (SEKF), it develops a direct radiance data assimilation system for the Common Land Model (CoLM). Assimilation experiments are conducted using brightness temperature data from the Microwave Radiation Imager (MWRI) onboard the FY-3D satellite. Their performance over China’s land areas is systematically assessed through comparison with the assimilation scheme based on the Community Microwave Emission Model (CMEM). The results show that the MLP-based assimilation scheme can effectively improve soil moisture simulation accuracy, yet the improvement varies across vegetation types: grassland areas achieve the largest error reduction (10.2%), while semidesert areas present the most prominent increase in the correlation coefficient (53.9%). Compared with the CMEM scheme, the MLP scheme exhibits better error stability and produces generally improved assimilation effects; specifically, in semidesert areas, the error decreases by 9.4%, and the correlation coefficient increases by 62.8%. This study demonstrates that deep learning-based observation operators have strong application potential for land surface data assimilation under complex physical mechanisms.