Large-scale hydrological studies are often limited by the lack of available observation data with a good spatiotemporal coverage. This has affected the reproducibility of previous studies and the potential improvement of existing hydrological models. In addition to the observation data themselves, insufficient or poor-quality metadata have also discouraged researchers from integrating the already-available datasets. Therefore, improving both the availability and quality of open water quality data would increase the potential to implement predictive modeling on a global scale. The Global River Water Quality Archive (GRQA) aims to contribute to improving water quality data coverage by aggregating and harmonizing five national, continental and global datasets: CESI (Canadian Environmental Sustainability Indicators program), GEMStat (Global Freshwater Quality Database), GLORICH (GLObal RIver CHemistry), Waterbase and WQP (Water Quality Portal). The GRQA compilation involved converting observation data from the five sources into a common format and harmonizing the corresponding metadata, flagging outliers, calculating time series characteristics and detecting duplicate observations from sources with a spatial overlap. The final dataset extends the spatial and temporal coverage of previously available water quality data and contains 42 parameters and over 17 million measurements around the globe covering the 1898–2020 time period. Metadata in the form of statistical tables, maps and figures are provided along with observation time series. The GRQA dataset, supplementary metadata and figures are available for download on the DataCite- and OpenAIRE-enabled Zenodo repository at https://doi.org/10.5281/zenodo.5097436 (Virro et al., 2021).

Persistent forest loss in the Brazilian Legal Amazon (BLA) is responsible for carbon emission, reduction of ecosystem services, and loss of biodiversity. Combining spatial data analysis with high spatial resolution data for forest cover and forest loss, we quantified the spatial and temporal patterns of forest dynamics in the BLA. We identified an alarming trend of increasing deforestation, with especially high rates in 2016 and 2017. Moreover, the creation of forest cover fragments is faster than ever due to decreasing size and dispersion of forest loss patches. From 2001 to 2017, the number of large forest loss patches decreased significantly, accompanied by a reduction in the size of these patches. Enforcement of field inspections and of initiatives to promote forest conservation will be required to stop this trend.

Nutrient runoff from agricultural production is one of the main causes of water quality deterioration in river systems and coastal waters. Water quality modeling can be used for gaining insight into water quality issues in order to implement effective mitigation efforts. Process-based nutrient models are very complex, requiring a lot of input parameters and computationally expensive calibration. Recently, ML approaches have shown to achieve an accuracy comparable to the process-based models and even outperform them when describing nonlinear relationships. We used observations from 242 Estonian catchments, amounting to 469 yearly TN and 470 TP measurements covering the period 2016-2020 to train random forest (RF) models for predicting annual N and P concentrations. We used a total of 82 predictor variables, including land cover, soil, climate and topography parameters and applied a feature selection strategy to reduce the number of dependent features in the models. The SHAP method was used for deriving the most relevant predictors. The performance of our models is comparable to previous process-based models used in the Baltic region with the TN and TP model having an R² score of 0.83 and 0.52, respectively. However, as input data used in our models is easier to obtain, the models offer superior applicability in areas, where data availability is insufficient for process-based approaches. Therefore, the models enable to give a robust estimation for nutrient losses at national level and allows to capture the spatial variability of the nutrient runoff which in turn enables to provide decision-making support for regional water management plans.

A Discrete Global Grid System (DGGS) is a type of spatial reference system that tessellates the globe into many individual, evenly spaced, and well-aligned cells to encode location and, thus, can serve as a basis for data cube construction. This facilitates integration and aggregation of multi-resolution data from various sources to rapidly calculate spatial statistics. We calculated normalized area and compactness for cell geometries from 5 open-source DGGS implementations - Uber H3, Google S2, RiskAware OpenEAGGR, rHEALPix by Landcare Research New Zealand, and DGGRID by Southern Oregon University - to evaluate their suitability for a global-level statistical data cube. We conclude that the rHEALPix and OpenEAGGR and DGGRID ISEA-based DGGS definitions are most suitable for global statistics because they have the strongest guarantee of equal area preservation - where each cell covers almost exactly the same area on the globe. Uber H3 has the smallest shape distortions, but Uber H3 and Google S2 have the largest variations in cell area. However, they provide more mature software library functionalities. DGGRID provides excellent functionality to construct grids with desired geometric properties but as the only implementation does not provide functions for traversal and navigation within a grid after its construction.

To understand, model, and predict landscape evolution, ecosystem services, and hydrological processes, the availability of detailed observation-based soil data is extremely valuable. For the EstSoil-EH dataset, we synthesized more than 20 eco-hydrological variables on soil, topography, and land use for Estonia (https://doi.org/10.5281/zenodo.3473289, Kmoch et al., 2019a) as numerical and categorical values from the original Soil Map of Estonia, the Estonian 5 m lidar DEM, Estonian Topographic Database, and EU-SoilHydroGrids layers. The Soil Map of Estonia maps more than 750 000 soil units throughout Estonia at a scale of 1:10 000 and forms the basis for EstSoil-EH. It is the most detailed and information-rich dataset for soils in Estonia, with 75 % of mapped units smaller than 4.0 ha, based on Soviet-era field mapping. For each soil unit, it describes the soil type (i.e. soil reference group), soil texture, and layer information with a composite text code, which comprises not only the actual texture class, but also classifiers for rock content, peat soils, distinct compositional layers, and their depths. To use these as eco-hydrological process properties in modelling applications we translated the text codes into numbers. The derived parameters include soil layering, soil texture (clay, silt, and sand contents), coarse fragments, and rock content of the soil layers within the soil profiles. In addition, we aggregated and predicted physical variables related to water and carbon cycles (bulk density, hydraulic conductivity, organic carbon content, available water capacity). The methodology and dataset developed will be an important resource for the Baltic region, but possibly also for all other regions where detailed field-based soil mapping data are available. Countries like Lithuania and Latvia have similar historical soil records from the Soviet era that could be turned into value-added datasets such as the one we developed for Estonia.

This study introduces a time-lag-informed Random Forest (RF) framework for streamflow time-series prediction across diverse catchments and compares its results against SWAT predictions. We found strong evidence of RF’s better performance by adding historical flows and time-lags for meteorological values over using only actual meteorological values. On a daily scale, RF demonstrated robust performance (Nash–Sutcliffe efficiency [NSE] > 0.5), whereas SWAT generally yielded unsatisfactory results (NSE < 0.5) and tended to overestimate daily streamflow by up to 27% (PBIAS). However, SWAT provided better monthly predictions, particularly in catchments with irregular flow patterns. Although both models faced challenges in predicting peak flows in snow-influenced catchments, RF outperformed SWAT in an arid catchment. RF also exhibited a notable advantage over SWAT in terms of computational efficiency. Overall, RF is a good choice for daily predictions with limited data, whereas SWAT is preferable for monthly predictions and understanding hydrological processes in depth.