Deep learning for epistemic uncertainty in SMAP-derived soil moisture estimates over the Kulfo watershed, Ethiopia

Precise soil moisture estimation is critical for irrigation scheduling and water resource management, especially in semi-arid regions like the Kulfo watershed, Ethiopia, where water availability is highly climate-dependent. However, quantifying the reliability of soil moisture estimates remains a key challenge. Standard satellite products like SMAP provide a measure of aleatoric uncertainty, but this offers limited insight into the confidence of a predictive model. This study develops a deep learning framework to produce a more reliable and better-calibrated measure of epistemic uncertainty, which directly quantifies analytical confidence. Using data from the Soil Moisture Active Passive (SMAP) mission, time-series data were processed in Google Earth Engine (GEE) and structured for deep learning models. Ten different models were tested: one basic long short-term memory (LSTM) model that doesn't consider uncertainty and nine uncertainty-aware model (including five deep ensemble models, Monte Carlo Dropout (MC Dropout), and Quantile Regression models). The Deep Ensemble model achieved the highest accuracy (RMSE=0.131, R 2 =0.993) and a 94.51% more reliable uncertainty estimate than the baseline. The MC Dropout model also delivered a much-improved confidence estimate, showing a 41.79% enhancement. Among the quantile regression models, the 5th percentile model produced a 40.58% better confidence estimate. Among the quantile regression models, the 5th percentile model produced a 40.58% better confidence estimate. This study demonstrates that coupling an LSTM-based model with the Deep Ensemble method provides a highly reliable and precise measure of model-specific analytical confidence for soil moisture estimates, offering a more trustworthy approach for decision-making in data-limited regions. • Deep learning quantifies epistemic uncertainty in SMAP soil moisture. • Deep Ensemble LSTM achieved highest accuracy & most reliable uncertainty estimates. • MC Dropout improved model confidence by 41.8%, Quantile Regression by 40.6%. • Framework enables more trustworthy soil moisture predictions in Kulfo watershed. • Uncertainty-aware models enhance decision-making in data-limited semi-arid regions.