How good are sediment measurements? An integrated approach to quantifying total uncertainty in metadata-limited annual suspended sediment yield observations

Deriving sediment yield (SY) from discharge (Q) and suspended sediment concentration (SSC) measurements is subject to multiple sources of error, including the sampling method, the sampling scheme and frequency, the load calculation method, and the measuring period. While the uncertainty from these individual sources of error has been studied in various contexts, their combined effect on SY calculations remains largely unquantified. This is mainly due to their complex and counteracting influences, the absence of detailed sampling protocol information, and the lack of true reference data. Still, estimating the total uncertainty on current and historical SY measurements is crucial to understand how these observational errors propagate in SY modelling and can impact subsequent model interpretation and decision-making.

Here, we aim to develop a tool that can provide realistic ranges of total uncertainty in SY observations worldwide for which limited metadata is reported. We do this by means of Monte Carlo simulations and machine learning. Using available long-term daily Q and SSC series, we quantify the effect of measurement-related sources of error, as well as their relative importance, on SY calculations. We apply this method on a (spatially) diverse selection of ∼180 gauging stations and further explore the relationship between SY uncertainty and catchment characteristics, including upstream area, land cover, and climate. Preliminary findings indicate that the range of uncertainty in SY calculations is mainly influenced by the sampling frequency, whereas the load calculation method and the sampling scheme can introduce important biases. Measuring errors on individual Q and SSC observations have relatively little impact on total SY uncertainty, provided that these measurements are unbiased. When considering long-term average SY, the length of the measuring period then becomes the most important source of uncertainty. Overall, the combined effect of these sources of error can lead to deviations up to three orders of magnitude from the true SY. Using these sampling-related variables and catchment characteristics derived from global hydro-environmental datasets, we further apply a gradient boosting algorithm to predict total uncertainty in annual SY and achieve a Nash-Sutcliffe model efficiency of ∼0.76. The model resulting from this work can thus provide scientists with realistic uncertainty estimates on existing SY observations with only basic metadata information available.