Land use and land cover changes and the terrestrial carbon sink are two important components of the global carbon budget. Several methodological approaches exist to measure fluxes of CO2 and other greenhouse gases between ecosystems and the atmosphere. With an accurate quantification of these fluxes, it is possible to compare carbon source and sink strengths between different land covers and to evaluate environmental influences on these terms. Out of those methods, the eddy covariance technique has the advantage of providing direct and quasi-continuous turbulent flux observations at the ecosystem scale. However, to compare eddy covariance data to, e.g., top-down methods and to achieve spatially gapless data sets, these point measurements with a relatively small footprint require a spatial upscaling with statistical methods such as machine learning and ancillary remote sensing data. Another issue with eddy covariance data sets is the underrepresentation of certain ecosystem types and climatic regions. Recently disturbed ecosystems belong to this group, but usually also exhibit non ideal characteristics for eddy covariance measurements such as abrupt surface changes and heterogeneous regrowth. Therefore, it is important to assess the uncertainty of eddy covariance measurements for disturbed ecosystems in regard to different choices of measurement design and processing and thus to improve the interpretability of such measurements. On the other hand, a changing climate can also enforce a reduced sink strength on ecosystems through, e.g., heat and drought. In this way, eddy covariance derived data on CO2 uptake in combination with other environmental measurements and advanced statistical analyses can reveal limiting conditions for photosynthesis and thus a reduced efficiency to use light for CO2 assimilation. In this dissertation, these three issues, i) spatial upscaling of eddy covariance data, ii) methodological uncertainties of obtaining flux data at disturbed sites, and iii) environmental impacts on ecosystem-scale photosynthesis, are addressed within the TERENO Eifel/Lower Rhine Valley Observatory, which comprises the Rur catchment, mostly located in western Germany. In a first study, eddy covariance CO2 flux data from different land covers within the Rur catchment were upscaled to the whole catchment area using a random forest machine learning model incorporating MODIS remote sensing and COSMO-REA6 reanalysis data. For this task, state-of-the-art predictor variable selection methods for machine learning models were evaluated. Results of this studys how that combining eddy covariance flux data with remote sensing products and reanalysis data is a feasible way to upscale CO2 flux information to the regional scale at a relatively high spatial resolution (250 m) and across various land covers. The study further indicates that averaging multiple model runs in the feature selection process can improve these results. Although an R² of 0.41 is in the range of other studies using a spatial cross validation scheme, this value reveals that there is still room for improvement. Main limitations of the analysis include a low prediction performance on high magnitude fluxes as a narrower range was predicted than observed, and the fact that differences between land cover classes were also narrower in the upscaled product than between eddy covariance stations. The further analyses were confined to a subregion within the Rur catchment, the Wüstebach site in the northern Eifel low mountain range. The site encompasses the Wüstebach headwater region and is mostly composed of a planted spruce forest but also contains a deforested area of 8.6 ha with unmanaged regrowth. This fast-growing vegetation requires a regular adjustment of the eddy covariance measurement height in order to ensure a stable flux source area in the long run and to prevent high spectral losses. In a second study, CO2 and H2O fluxes were hence measured over the deforested area with eddy covariance systems in two different heights and were processed with five different spectral corrections. In this way, the uncertainty from measurement height and choice of spectral correction was assessed, and insights were gained in the trade-offs that must be considered at a site with non-ideal characteristics. For the deforested site, results show that at the lower height spectral corrections were higher and had a higher standard deviation among methods compared to the upper height for both CO2 and H2O fluxes. The average standard deviation between heights was even higher than between spectral corrections at the same height (24.8% of CO2 flux; 9.7% of H2O flux). Furthermore, the energy balance closure was on average about 9% better for the upper system than for the lower system. On the other hand, the modelled footprints of both heights did not match the average footprint of the previous years at the lower height. Hence, the study indicates a difficulty of achieving a stable flux source area over longer time periods for fast growing vegetation but also emphasizes the importance of a carefully adjusted measurement height. Although the study improved the interpretability of flux measurements for a disturbed site, its main limitation comprises the difficulty to apply one of the common footprint models to estimate the flux source area for this site with complex flow, especially over the forest edges. A third study concerned the Wüstebach spruce forest. For this site gross primary productivity derived from eddy covariance CO2 flux data was combined with measurements of green canopy absorbed photosynthetically active radiation (APARg), sap flow, and other meteorological and plant physiological data. In this way, water-limiting conditions for photosynthesis and the light use efficiency of a spruce forest were evaluated. In addition, the importance of environmental variables for the prediction of gross primary productivity was assessed with state-of-the-art machine learning variable importance measures. In this study, data from the 2021 growing season was analyzed, for which the light use efficiency of green parts of the forest was on average 4.0 ± 2.3% and showed a unimodal relation to air temperature with a maximum around 15 °C. For modelling gross primary productivity with tree based machine learning models, canopy chlorophyll content likely as a seasonal variable for photosynthetic capacity and APARg likely as a diurnal variable for energy supply were the most important variables. On days with high vapor pressure deficit, tree-scale sap flow and ecosystem-scale gross primary productivity both shifted to a clockwise hysteretic response to APARg. It is demonstrated that the onset of such a clockwise hysteretic pattern of sap flow to APARg can be a useful indicator of afternoon stomatal closure related to water-limiting conditions. However, the main limitation of this case study is its limited extent, as just one comparatively cool and wet growing season at a single site with a single dominant tree species, Picea abies, was investigated. Overall, this dissertation highlights the use of direct flux measurements and machine learning methods for both the evaluation of land cover changes and the impact of changing environmental conditions on the CO2 source and sink strengths of terrestrial ecosystems.