High-resolution European daily soil moisture derived with machine learning (2003-2020)

Area of Science:

• Earth Science
• Environmental Science
• Data Science

Background:

• Machine learning (ML) offers a novel approach for generating large-scale land surface data.
• ML models can estimate soil moisture by learning relationships between meteorological variables and in-situ measurements.
• Physics-based knowledge is not required for ML-driven soil moisture estimation.

Purpose of the Study:

• To develop a high-resolution (0.1°) daily soil moisture dataset for Europe (SoMo.ml-EU).
• To improve the accuracy and utility of soil moisture data for regional studies.
• To provide an observation-based dataset for hydrological and agricultural analyses.

Main Methods:

• Utilized Long Short-Term Memory (LSTM) networks for model training.
• Trained the model using in-situ soil moisture measurements.
• Developed a dataset covering three vertical layers for the period 2003-2020.

Main Results:

• The SoMo.ml-EU dataset shows closer agreement with independent in-situ data compared to previous lower-resolution versions.
• High-resolution meteorological data combined with in-situ observations enhances dataset accuracy.
• The dataset effectively describes soil moisture variability, including drought conditions, when compared regionally.

Conclusions:

• The new SoMo.ml-EU dataset provides enhanced, high-resolution, observation-based soil moisture data for Europe.
• This dataset is valuable for regional hydrological and agricultural studies.
• The joint processing of in-situ and high-resolution meteorological data improves soil moisture estimation.