Nicholas James Gubbins , Weston M Slaughter, Michael Vlah , Spencer Rhea, William H McDowell, Matthew RV Ross, Emily Bernhardt

Streams and rivers export dissolved materials and eroded sediments from the watersheds they drain. Much can be learned about rivers and their watersheds by measuring the magnitude, timing and form of these exports. Such watershed load datasets are used to gain fundamental understanding of watershed ecosystems as well as to assess water quality and the efficacy of management approaches to sustain both terrestrial and aquatic ecosystem health. Despite the widespread use of watershed load estimates, comparisons at macroscales (i.e. across many sites) are currently complicated by differences in underlying data quality and estimation methods between sites, periods, and solutes. Using high-frequency sensor data from the Hubbard Brook Experimental Forest and the Plynlimon Research Catchments, we generated time series of increasingly coarse sampling frequencies, and tested the sensitivity of various load estimation methods. We further tested the accuracy of common methods using synthetic time series, spanning a range of flow regimes and concentration-discharge (C:Q) relationships. Lastly, we applied each estimation method to the MacroSheds dataset (macrosheds.org), generating a publicly available dataset of 16,489 site-years of data across 93 sites and 112 solutes. Results from both the simulated data coarsening and synthetic time series experiments indicate that load estimates with high sampling frequency (daily or better) and an informative concentration-discharge relationship are well suited for macroscale science efforts (errors within ~10%). Estimates based on coarse (biweekly or coarser) underlying data and incompletely described and/or complex C:Q relationships showed large enough error (>50%) to suggest they would be misleading if included in macroscale efforts. Our results suggest that scientists interested in comparing load estimates should first consider (1) sensitivity of their analysis to changes in load magnitudes, (2) the underlying data frequency used to generate estimates, (3) the C:Q relationship of their solute of interest, and (4) their confidence in the completeness of that C:Q relationship over the period of study.