Global soil moisture data derived through machine learning trained with in-situ measurements

Area of Science:

• Earth Science
• Hydrology
• Climate Science

Background:

• Spatially continuous soil moisture data is crucial for hydrology and climate studies but often limited to satellite or model outputs.
• Existing datasets may not fully capture the complex spatial and temporal dynamics of soil moisture.

Purpose of the Study:

• To introduce SoMo.ml, a novel global, long-term soil moisture dataset.
• To provide a machine learning-derived dataset that complements existing soil moisture data sources.

Main Methods:

• Utilized a Long Short-Term Memory (LSTM) model trained on in-situ measurements from over 1,000 global stations.
• Extrapolated daily soil moisture dynamics across space and time.
• Generated multi-layer soil moisture data (0-10cm, 10-30cm, 30-50cm) at 0.25° spatial resolution for 2000-2019.

Main Results:

• SoMo.ml provides a global, long-term record of multi-layer soil moisture.
• The dataset demonstrates strong performance in capturing temporal soil moisture dynamics.
• Cross-validation and inter-comparison confirm the dataset's reliability.

Conclusions:

• SoMo.ml offers a valuable, distinct data source for large-scale hydrological, meteorological, and ecological analyses.
• Its strength in temporal dynamics makes it ideal for applications like anomaly detection and memory analyses.
• This dataset enhances the suite of available tools for understanding Earth's systems.