WATERLINE

Hydrological models are vital tools for water resources management. Advanced computational algorithms can simulate the relevant physical processes and form the feedback mechanism across a wide range of spatial and temporal scales. However, a bottleneck of these models is the lack of environmental observations to calibrate model parameters and to assess the robustness of model predictions. It is the objective of the WATERLINE project to improve hydrological models and their predictions based on observations from different sources, such as satellites, drones and meteorological stations. For this purpose, observations from different climatic regions in Europe will provide measurement data on important hydrological variables, e.g. precipitation and soil moisture. In addition, citizens and institutions from the education sector will be engaged via citizen science and provide observations and measurements to the project data pool. In order to bridge the gap between the often coarse-scale satellite data and the requirements of users and stakeholders at the small scale, novel downscaling approaches will be developed including the use of machine learning. These downscaled fields of precipitation and soil moisture will be used to force hydrological models to obtain high-resolution (in terms of both space and time) predictions of hydrological parameters, such as groundwater levels, river discharge and evapotranspiration. Such predictions will be applied in water resources management, in the agricultural sector and in the educational sector. In collaboration with users and decision makers from these sectors the project partners will develop novel visualisation methods in order to maximise the use and uptake of the project results. The WATERLINE consortium combines the expertise of hydrologists, computer scientists and experts in earth observation and consists of partner organisations from Greece, Austria, Finland, Poland, and Switzerland.

Hydrological models are essential tools in water resources assessment and management. Advanced computational algorithms can simulate the relevant physical processes and form the feedback mechanism across a wide range of spatial and temporal scales. However, a bottleneck of these models is the lack of environmental observations to calibrate model parameters and to assess the robustness of model predictions. Unfortunately, neither in-situ networks nor remote sensing alone can provide sufficient information to capture the high spatial and temporal variability of hydrological processes. Recently, downscaling frameworks have been developed, building robust models between coarse scale products and high-resolution covariates using in-situ measurements. WATERLINE employs multi-source information from remote sensing, historical data, in-situ data from meteorological networks as well as crowdsourced measurements to improve hydrological models and their predictions. TU Wien has developed a novel downscaling framework for satellite soil moisture retrievals in this context. This approach employs a two-stage machine learning model, which at the core is based on a random forest regression algorithm. Coarse resolution soil moisture at 0.25 degree resolution from the European Space Agency (ESA) Climate Change Initiative (CCI) is the main input. These data are based on harmonized multi-satellite retrievals, which allow the creation of a global, long-term record for climate change assessments. The downscaling framework also uses soil moisture covariates to establish the required models to go from coarse to medium (5 km) and subsequently to high resolution (1 km). The predictors use a mix of dynamic (vegetation indices) and static (land cover and topography, soil properties) information in this two-stage approach. This assumes that the statistical relationships between the ancillary variables and SM is not significantly affected by the spatial scale. The downscaled soil moisture dataset was evaluated using the original ESA CCI dataset, with data from hundreds of ground soil moisture measurements from traditional sensors, and from crowd-sourced data. As part of this evaluation, the new record was also compared to other high-resolution soil moisture products, such as a downscaled product based on NASA's SMAP satellite as well as native high-resolution soil moisture from Sentinel-1. In all comparisons, it was found that the temporal dynamics of in situ soil moisture were well represented by the downscaled product. Comparison to the native data records shows that downscaled product tends to contain less noise (estimated using Triple Collocation Analysis). The dataset was produced over Europe for the last 17 years (2008-2024). Soil moisture data at high spatial resolution enables new applications such as irrigation detection and agricultural drought assessment. At the same time, downscaling approaches can be applied to existing long-term soil moisture records, which enables the creation of multi-year or even multi-decade records, that can adequately consider the impact of climate change.