Sediment fingerprinting is widely used in drainage basin analysis to identify the provenance and source contributions of sediments (or other material) in transit from source-to-sink. By enabling source areas of sediment supply to be targeted, the method has become an integral part of sustainable landscape management. The precision and accuracy of sediment fingerprinting is contingent on the choice of mixing model, which quantifies the contribution of potential sediment sources by minimizing the difference between observed properties of sink samples and characteristic properties of the sources. Here, we apply a set of frequentist and Bayesian mixing models with the aim of identifying the optimum composite fingerprint of four sediment sources (viz., agricultural land, rangeland, gullies, and landslides) in a small catchment draining the Iranian Loess Plateau in the Golestan province of northeastern Iran. Forty-four soil samples were collected from the four potential source zones. Based on seven synthetic mixtures with known source contributions we compared the performance of a frequentist Monte Carlo model, GLUE model, a Bayesian end-member model (BEMMA), MixSIAR Bayesian model, and a Brewer Bayesian model. We found that, in terms of uncertainty estimation, the best results were obtained with GLUE and BEMMA. Applying GLUE to our study catchment, we estimated the following source contributions to an earth dam reservoir: agricultural land (55.8 %), rangeland (33.7 %), gullies (15.7 %), and landslides (14.2 %), confirming the view that agriculture is the main cause of reservoir sedimentation. All source contributions exhibited high variability, which we attribute to storm frequency, sediment delivery due to hillslope-sink connectivity, and human activities involving removal of vegetation.