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

Sungmin O,Rene Orth,Ulrich Weber,Seon Ki Park
DOI: https://doi.org/10.48550/arXiv.2205.10753
2022-05-22
Abstract:Machine learning (ML) has emerged as a novel tool for generating large-scale land surface data in recent years. ML can learn the relationship between input and target, e.g. meteorological variables and in-situ soil moisture, and then estimate soil moisture across space and time, independently of prior physics-based knowledge. Here we develop a high-resolution (0.1°) daily soil moisture dataset in Europe (<a class="link-external link-http" href="http://SoMo.ml" rel="external noopener nofollow">this http URL</a>-EU) using Long Short-Term Memory trained with in-situ measurements. The resulting dataset covers three vertical layers and the period 2003-2020. Compared to its previous version with a lower spatial resolution (0.25°), it shows a closer agreement with independent in-situ data in terms of temporal variation, demonstrating the enhanced usefulness of in-situ observations when processed jointly with high-resolution meteorological data. Regional comparison with other gridded datasets also demonstrates the ability of <a class="link-external link-http" href="http://SoMo.ml" rel="external noopener nofollow">this http URL</a>-EU in describing the variability of soil moisture, including drought conditions. As a result, our new dataset will benefit regional studies requiring high-resolution observation-based soil moisture, such as hydrological and agricultural analyses. The <a class="link-external link-http" href="http://SoMo.ml" rel="external noopener nofollow">this http URL</a>-EU is available at figshare.
Atmospheric and Oceanic Physics
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the challenge of obtaining high - resolution soil moisture data. Specifically, although there are various soil moisture data sources, it is still very difficult to obtain high - resolution (for example, 0.1°) soil moisture data. The existing large - scale soil moisture data sets are usually of coarse resolution (25 - 50 kilometers), which limits the consideration of surface heterogeneity and related physical processes (such as land - atmosphere coupling). In order to meet the growing demand for high - resolution data, this study uses machine - learning techniques, especially the Long Short - Term Memory network (LSTM), combined with ground - based observation data and high - resolution meteorological data, to generate a new daily soil moisture data set for Europe (SoMo.ml - EU). This data set covers the period from 2003 to 2020 and provides daily soil moisture information at multiple layers (0 - 10 cm, 10 - 30 cm, 30 - 50 cm) with a spatial resolution of 0.1°. Through this method, the study aims to provide a more detailed data set that more accurately reflects soil moisture dynamics to support regional research and applications that require high - resolution soil moisture information, such as hydrological and agricultural analysis.