Satellite microwave radiometers effectively monitor landscape freeze/thaw (FT) transitions but have difficulty distinguishing soil from other landscape properties, which can lower retrieval accuracy. Here, we applied a deep learning model for soil FT classification driven by daily brightness temperatures (TBs) from AMSR2 and SMAP, and trained on soil (~0-5cm depth) FT observations. The probability of frozen or thawed conditions was derived using a model cost function optimized using observational training data over the Northern Hemisphere (NH) and five year (2016-2020) study period. Results showed favorable accuracy against soil FT observations from ERA5 reanalysis (mean annual accuracy, MAE: 92.7%) and NH weather stations (MAE: 91.0%). Moreover, SMAP L-band (1.41 GHz) TBs provided enhanced soil FT performance over alternative retrievals derived using only AMSR2 inputs. FT accuracy was also consistent across different land covers and seasons. The results provide better soil FT precision to improve understanding of complex seasonal transitions and their influence on ecological processes and climate feedbacks.