Prediction-Constrained Multimodal Temporal Fusion of Satellite Microwave and Infrared Observations for Continuous Snow Temperature Profile Estimation

Snow temperature profiles are important for characterizing snowpack thermal state and vertical heat transfer, but continuous estimation is limited by sparse in situ observations and discrete satellite overpasses. This study proposes a prediction-constrained multimodal temporal fusion method combining continuous deep-learning prediction with satellite microwave–infrared updates. The Weather-Forced Residual LSTM with Persistence Fallback (WFRL-P) generates an hourly multi-depth prediction background from historical snow temperatures and meteorological forcing. Snow temperature profiles are then conjointly retrieved from AMSR2 microwave and MODIS infrared brightness temperatures, and valid retrieval–prediction differences are transformed into depth-dependent, magnitude-limited updates and causally propagated for up to 6 h. Independent evaluation at Dome C during 2020–2021 produced hourly estimates at 12 depths from 5 to 1000 cm, with an all-depth bias of −0.1263 K, MAE of 0.3657 K, RMSE of 0.8218 K, and R2 of 0.9880. At most comparable depths, the estimation errors were lower than those reported in previous Antarctic snow temperature remote-sensing studies. Multimodal fusion further reduced the all-depth MAE of the fixed lead-24 prediction background by 3.32%, with a 5.14% improvement at matched satellite-overpass samples. The results demonstrate that satellite microwave and infrared observations can provide effective incremental constraints while maintaining the overall accuracy of the continuous prediction background, offering a feasible approach for the temporal fusion of discrete satellite observations and continuous snow temperature predictions.