A deep learning emulator for scalable soil moisture monitoring based on satellite and land data assimilation system

Soil moisture is fundamentally important for drought monitoring, hydrologic forecasting, weather and climate predictions, agriculture and forest management, and many other applications. Advancements in satellite observations and land data assimilation systems (LDAS) have created new opportunities for field-scale soil moisture monitoring locally and across the globe. Using satellite and LDAS data to estimate soil moisture across multiple scales would benefit many applications and scientific research, especially for locations where ground soil moisture observations are not available. In this study, we introduce a deep learning emulator, namely Scalable Deep Learning for Soil Moisture Monitoring (SDLS), for root-zone soil moisture monitoring from the field to the regional scales. The SDLS method uses LDAS forcings and simulations from sampled locations to a bidirectional long short-term memory (B-LSTM) deep learning model, and is further applied to 30-m satellite-based evapotranspiration (ET), land cover, and topographic data, and digital soil property data. Evaluation of SDLS emulations demonstrates robust performance, with a mean squared error (MSE) below 0.0004, a Pearson correlation coefficient exceeding 0.8, and a Kling-Gupta Efficiency (KGE) score above 0.75 against LDAS soil moisture. SDLS method can generate daily soil moisture at 30-m resolution and can capture field-scale variability and drought, well matching with in situ observations. With additional deep learning postprocessing, the performance of the SDLS soil moisture against in situ observations can be further improved. The strength of the SDLS method lies in its ability to leverage process-based physical knowledge in land surface models to estimate soil moisture using satellite observations in a scalable way, which can be readily applied to new locations without the need for ground observations.