Observation-based reconstruction of riverine organic carbon fluxes reveals hydroclimatic controls on lateral carbon export

Lateral export of organic carbon by rivers links terrestrial and aquatic carbon cycling, yet its magnitude, drivers, and variability remain poorly quantified at large spatial and temporal scales. Here we develop a physics-constrained multi-task machine learning model and use long-term in situ riverine total and dissolved organic carbon (TOC/DOC) observations to reconstruct daily TOC concentrations and fluxes at 0.25° resolution across the contiguous United States (CONUS) for the past four decades (1984–2023). The multi-task learning approach leverages DOC-rich records to inform TOC dynamics through their observed covariation, improving TOC estimates in regions with sparse measurements, particularly in the arid western United States. The reconstructed data reveal a widespread decoupling between TOC concentrations and fluxes, with concentration trends increasing over 47% of the domain while fluxes decline over 73%, indicating a dominant role of hydroclimatic control on transport efficiency rather than changes in carbon source availability alone. Analysis across dry and wet years shows that wetter hydroclimatic conditions, particularly following drought periods, are associated with pronounced TOC export, during which lateral carbon export can exceed 10% of concurrent terrestrial carbon uptake. These results demonstrate how hydroclimatic variability modulates organic carbon transport in river networks, with implications for estimating land carbon storage and land-water coupling under ongoing hydroclimatic change. We emphasize the importance of integrating large-sample, in situ riverine observations in improving understanding of coupled hydrological and biogeochemical processes from site to continental scales.