Comparison of Machine Learning Algorithms for Simulating Brightness Temperature Using Data from the Tianjun Soil Moisture Observation Network

The L-band radiative transfer-forward modeling plays a crucial role in data assimilation for meteorological forecasting. By utilizing information from the underlying surface (typically land surface parameters and variables), such as soil moisture, soil temperature, snow cover, freeze–thaw status, and vegetation, the corresponding brightness temperatures can be simulated through the physical processes described by radiative transfer models. Data assimilation becomes meaningful when the errors introduced by the simulated brightness temperatures are smaller than the simulation accuracy of the land surface variables. However, radiative transfer models at the L-band cannot accurately simulate TB operationally. In this study, four machine learning methods, including random forest (RF), long short-term memory (LSTM), support vector machine (SVM), and deep neural networks (DNN), are employed to reconstruct the forward relationship from land surface parameters to brightness temperatures, serving as an alternative to traditional radiative transfer models. The performance of these methods is evaluated using ground-truthed soil moisture data, soil texture static data, and leaf area index (LAI). The results indicate that DNN and RF exhibit superior performance, with DNN achieving the lowest average unbiased root mean square error (ubRMSE) of 6.238 K for vertical polarization brightness temperature (TBv) and 9.033 K for horizontal polarization brightness temperature (TBh). Regarding correlation coefficients between the retrieved brightness temperatures and satellite measurements, RF leads for H-polarized TB with a value of 0.943, while both RF and SVM perform well for V-polarized TB with values of 0.930 and 0.932, respectively. In conclusion, our study shows that DNN is the optimal method for retrieving brightness temperatures, outperforming other machine learning approaches regarding error metrics and correlation with satellite measurements. These findings highlight the potential of DNN in improving data assimilation processes in meteorological forecasting.