<![CDATA[Soil moisture is an important indicator in the management of water resources, precision agriculture, and disaster mitigation, such as drought and land fires. Fluctuations in soil moisture are influenced by various climate variables, requiring a reliable predictive approach essential. This research develops a daily soil moisture prediction model using Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) algorithms with univariate and multivariate approaches. Soil moisture data were obtained from Google Earth Engine, while climate data were collected from 10 BMKG stations in East Java for the period 2019–2024. Data preprocessing includes cubic spline interpolation to handle missing values and Min-Max normalization to achieve uniform feature scaling. Models were built using a direct forecasting approach for horizons to and five evaluation metrics: MAE, MSE, RMSE, MAPE, and R². The results show that the multivariate GRU model performs best at horizon with MAE = 0.05455, MSE = 0.00604, RMSE = 0.07539, MAPE = 0.19280, and R² = starting from 0.9626 on day 1 (t), then decreasing to 0.8075 on day 10 The univariate LSTM model excelled in training time efficiency (<400 seconds) at most stations. The multivariate GRU model demonstrates the highest accuracy and stability, making it suitable for medium- to long-term forecasting, while the univariate LSTM excels in training speed, making it effective for daily predictions. The model’s performance remains limited to the dataset's spatial and temporal scope. Therefore, future research should test the model in other regions and under extreme climate conditions, as well as apply transfer learning in data-scarce areas. The novelty of this study lies in comparing LSTM and GRU performance for daily soil moisture prediction in both univariate and multivariate scenarios, using complete climate variables from multiple stations.]]>
