Data-driven soil moisture estimations based on earth observation data and machine learning
8 February 2022, 13:00–16:00, Zoom
Although its origins lie in military applications, satellite remote sensing was established over the last decades as an essential method for observing our environment through the spatially continuous and global measurement of relevant parameters like land-cover, land-surface temperature, vegetation biomass, ocean salinity, or surface soil moisture. Furthermore, it plays an essential role in observing the atmosphere, mapping land-cover and land-use changes, generating digital elevation models, and many other applications.This thesis focuses on estimating the surface soil moisture content, which was recognised as an essential climate variable by the Global Climate Observing System of the World Meteorological Organization. It is essential for the understanding of many meteorological and hydrological processes. Spatial and temporal changes of the soil moisture content can help understand and anticipate natural hazards like landslides, floods, or drought. The remote sensing of the soil moisture content, using optical and microwave sensors, has a long history dating back to the 1970s. Different approaches have emerged and established themselves since then. With this thesis, we concentrated on the latest group of approaches, machine learning.Even though most of the underlying methodologies were already developed during the 1980s and 1990s, machine learning experienced a surge of popularity during the last decade, also for remote sensing and earth observation applications. This increase in popularity was further pushed by the paradigm shift in earth observation, which allows users today to easily access and exploit large quantities of data from different sensors.The overall goal of this thesis is to use earth observation data combined with machine learning methods to estimate the soil moisture content. To capture the aims of this thesis, we formulated three main questions: How can we go from site-specific, data-driven machine-learning models to general applicability in large scale applications?; How can we harness the potential of available data and merge data from different sensors and data sources with different spatial and temporal resolutions?; How can we link soil moisture measurements across scales (spatial and temporal)?The analysis presented by the thesis in its first part focused on a better understanding of the interactions between microwaves, soil moisture, topography, land-cover, and vegetation, between each other and across spatial scales. By developing a spatial upscaling method for in-situ measurements, we were able to study these interactions and confirm the solid temporal correlation of soil moisture across spatial scales. An essential ambition of this work was also to study the applicability of data-driven methods on a global scale. For this purpose, we performed tests based on different spatial resolutions and used different reference data types. The results demonstrate that accurate estimation is possible, with coarse as well as with high spatial resolutions. The studies also revealed certain limitations related to the potential of retrieval models relying only on satellite data, the uncertainties of heterogenous reference data, or the validation of high-resolution spatial patterns.One of the thesis’ main outputs is an approach and a model for the high-resolution mapping of surface soil moisture, which we published as part of a software called PYSMM. The practical use of the approach we demonstrated as part of the thesis for mapping soil moisture anomalies. Its relevance was further underlined as it was picked up by scientists at FAO and the USGS to incorporate soil moisture information for wetland detection and assimilation in a hydrological model, respectively.