Purpose The results of sediment source tracing in large river catchments depend on defined sources being adequately represented by the sampling and in subsequent numerical analysis. We hypothesise that surface soil concentrations of fallout radionuclides caesium-137 ( 137 Cs) and lead-210 excess ( 210 Pb ex ) are smaller at locations with higher soil erosion rate and that if this is not accounted for, then spatially random sampling gives a biased representation of surface soil delivered to rivers and biased source contribution estimates. Materials and methods Surface soil was sampled across the Burdekin River basin in northeast Australia at 90 locations stratified by three classes of modelled soil erosion rate and analysed by gamma spectroscopy. Separate probability distributions (density functions) were fitted to the 137 Cs concentrations of samples of each erosion class, of subsurface soil and of river sediment. Surface soil distributions were aggregated by weighting in proportion to the upstream area and mean erosion rate of each erosion class, so that the high erosion class contributed disproportionately to the tracer properties of the surface soil source. Source contributions were estimated using a Monte Carlo mixing model. Results and discussion The mean surface soil concentrations of 137 Cs and 210 Pb ex were significantly different between soil erosion classes as hypothesised. Weighting surface soil from the high erosion class more heavily increased the estimated proportion of river sediment contributed from surface soil, by 35 % larger than if surface soil sampling was confined to low erosion areas. Stratified sampling and weighting by erosion rate is of greater importance in river basins with large gradients in soil erosion and where surface soil contributes substantially to river sediment. Surface soil contributed 6 % to fine sediment at the basin outlet and 0-14 % in major tributaries, which was somewhat lower than in a prior study probably due to recent above-average rainfall increasing vegetation ground cover. Conclusions Surface soil sampling for source tracing using fallout radionuclides should be stratified by erosion rate. The tracer properties of high erosion areas should be weighted more heavily than low erosion areas in source mixing models. If comprehensive sampling cannot be afforded, then sampling should be biased towards more highly eroding areas. The approach should be considered for other source tracers whose properties may co-vary with soil erosion rate. Fine sediment delivered from the Burdekin River basin to the Great Barrier Reef lagoon in recent decades was predominantly derived from gully erosion, streambank erosion and rilled and scalded areas on hillslopes.