Water in the root zone links the terrestrial water, carbon, and energy cycles, but only the part that ecosystems can access and use controls vegetation's resilience to drought and its feedbacks to climate. This water cannot be observed directly, and existing soil-moisture products report the total root-zone water instead of the fraction ecosystems draw on. Although this component can be inferred in gauged basins by an ecohydrological model that inverts the water balance, these basins cover only a small part of the land surface, and therefore global estimates require generalizing what is learned there to a continuous field over the entire land. We introduce MOREDO (Multi-Observation Root-zone Ecohydrology DiagnOstic model), a differentiable hybrid framework that couples a conceptual ecohydrological model with a neural-network parameterization. The network learns mappings from static environmental attributes to model parameters, and the framework is trained end-to-end against observed evapotranspiration, runoff, snow water equivalent, and terrestrial water storage anomalies across more than one thousand basins. Because these mappings depend only on globally available attributes, the basin-trained model can be applied on a continuous global grid. In basins withheld from training, MOREDO reproduces these four targets, and across the much larger, unsampled grid it produces fluxes and storage dynamics that remain regionally and seasonally consistent. Overall, MOREDO diagnoses the ecosystem-accessible store and its actively used component across the global land surface, turning information-rich but spatially sparse basin observations into large-scale understanding of root-zone ecohydrological dynamics that no observation captures directly.