No description available.

Accurately modeling climate and the impacts of climate change relies heavily on extensive observations. Soil moisture is a critical variable in this regard, as it influences energy partitioning, regulates the water cycle, directly affects vegetation dynamics, modulates terrestrial carbon sinks and sources, and overall plays a vital role in the land-atmosphere interactions and feedback. This work aims to improve the quality of available surface soil moisture data and its complementary dataset -- vegetation optical depth (since both are derived from the same satellite measurements). The datasets developed in the scope of this study fill the gap in the available observational data pool as unique, long-term, consistent datasets developed based on remote sensing data. These datasets were created with the help of machine learning tools, in particular, deep dense neural networks. The distinctive characteristics of the utilized approach include (1) decomposition of the signal into seasonal and residual parts and training a neural network to match the residuals; (2) applying a special transfer learning training scheme that allows adjusting the features of a trained neural network to a slightly different input that ultimately permits merging the non-compatible directly and disjoint satellite sources into a consistent dataset; (3) using an ensemble of neural networks to assess the data uncertainty. Upon development, the datasets were profoundly validated vs. in-situ soil moisture measurements for soil moisture and biomass and photosynthesis-related datasets for vegetation optical depth. The consistent and long-term nature of the created datasets allowed for the study of decadal trends in soil moisture and the potential drivers for its dynamics. Finally, this study presents two showcases of the datasets used for constraining models -- as data assimilated in a simple carbon cycle model and as an emergent constraint in an ensemble of global climate models. The vegetation optical depth dataset was used in a simple carbon cycle model and demonstrated how it can constrain unobserved respiration flux and carbon pools. In this project's scope, the role of information content, data quality, and local conditions is assessed. The soil moisture dataset is used to constrain global climate models' projections of future soil moisture change by constraining the past soil moisture change range. Altogether, this study proposes a robust methodology for merging data from different sources into a consistent long-term dataset (provided that at least a short overlap in data exists for transfer learning). The analysis of the soil moisture dataset reveals that the regions of drying and wetting dynamics exist globally and can be identified with statistically significant trends in soil moisture. The dynamics are studied seasonally, revealing the contradicting trends in soil moisture in some regions (for example, in Europe, wetting in spring and drying in summer) and persistent trends throughout the year for others (for example, drying in the Mediterranean). Similarly, the local drivers of the soil moisture change are established. The soil moisture change is mainly driven by variations in precipitation for dry regions and in temperature in wet regions with the rising role of vegetation dynamics, especially in high latitudes. Finally, the vegetation optical depth data has proven its high potential in constraining respiration flux and carbon pools, significantly improving the carbon cycle model predictions in the regions subjected to interannual variability in meteorological forcing conditions and vegetation response.