Time series deep learning has been shown to be a powerful tool in harnessing newly available big data for hydrologic predictions. To be more useful in forecast and projection applications, we must also be able to estimate model uncertainty. However, uncertainty quantification techniques previously did not exist for hydrologic DL models. Here we adapt an efficient uncertainty estimation framework, which originated from vision deep learning, for hydrologic time series predictions. This method simultaneously estimates heteroscedastic aleatoric uncertainty (attributable to inherent measurement/input noise) and epistemic uncertainty (attributable to imperfect models and insufficient training data). The aleatoric uncertainty is estimated by tuning an observation noise parameter. The epistemic component was based on Monte Carlo dropout, a method that treats the dropout training as an approximate Bayesian inference in deep Gaussian processes. We show that for reproducing soil moisture dynamics recorded by the Soil Moisture Active Passive (SMAP) mission, most uncertainty arises from dense vegetated or cold regions where SMAP’s error is relatively large or there are less available SMAP data to train the model. As expected, the epistemic uncertainty can be reduced by changing the amount of training data but not the aleatoric. Spatial extrapolation tests show that the epistemic error can well encompass the error encountered for unfamiliar examples, while the estimates for aleatoric uncertainty can be flawed if training data is not representative. Noise augmentation experiments further show that SMAP error mainly arises from its measurement error rather than forcing data. As a result, our SMAP aleatoric uncertainty map can approximate an uncertainty map for SMAP over continental United States. This estimate previously did not exist.