Streams are key components of the global carbon (C) cycle. The river network conceptualization moved from the ``pipeline” idea, in which C is solely transported, to the ``active ecosystem” concept, in which both terrestrial derived and autochthonous C are processed and transported. Functioning, at ecosystem level, is encompassed by ``Ecosystem Metabolism” which measures production (GPP) and respiration (ER) of organic matter within a stream reach. Proper estimation of GPP and ER is relevant as it bridges the terrestrial and aquatic C cycles through lateral carbon fluxes, and informs on the biogeochemical connectivity downstream and throughout fluvial networks. Whole-ecosystem metabolism has received considerable attention over the last decades. However, even though ecosystem metabolism and biomass dynamics are tightly coupled, these two components have often been studied separately. Moreover, despite the natural link between ecosystem-level processes in flowing waters, and the inherent heterogeneity of all metabolic drivers at the catchment scale, no study has attempted to upscale GPP and ER from different and separated river reaches to an entire river network, continuously, both in time and space. In this perspective, this Thesis develops a multi-scale modeling framework for the investigation of riverine metabolism. In particular, the Thesis provides, firstly, a novel process based reach-scale model able to locally estimate metabolic fluxes based on measurements of dissolved oxygen (DO), by coupling DO dynamics with two non-observed state variables describing the temporal evolution of autotrophic and heterotrophic benthic biomass. The model integrates current knowledge on metabolic drivers and poses itself as a complementary tool to the widely employed single station approach. Furthermore, this Thesis illustrates the potential of machine learning in dealing with the upscale problem. Random forest has been used to extrapolate both in time and space reach-scale estimates of ecosystem metabolism to the scale of an entire stream network. This approach has been instrumental to establish annual regimes across an entire river network, to quantify the importance of metabolic drivers, to assess the relative contributions from small and large streams, and to disclose properties emerging from the network they form. Metabolic stability, allochthony, and scaling induced by the stream network have been investigated. Furthermore, the same algorithm has been applied to local observations of streamwater temperature and photosynthetic active radiation which made them available over the entire river network. This allowed us to couple the climatic regime with the reach-scale ecosystem model and integrate network structure, land cover, and the hydrological regime. We thus developed a process-based spatially-distributed framework able to simulate the ecological functioning of organic matter (and thus ecosystem metabolism) at network-scale, to investigate the variability with which ecosystems use available energy throughout the river network, and to directly assess the effect of climatic and hydrological change. We leveraged on data availability in the Ybbs river network (Austria), to evaluate the models' capability in reproducing local patterns of GPP and ER. The case study confirmed the reliability of the model frameworks and provided empirical evidence for long-standing theory predicting shifts of ecosystem metabolism along the stream continuum.