The backgrounds of stream sediment geochemical samples are associated with the underlying geological bodies. Moreover, a stream sediment geochemical data set is a closed number system because it contains compositional variables that are parts of a whole. Consequently, the empirical frequency distributions of stream sediment geochemical data are often skewed or with multiple peaks. While it is clear that data should approach a symmetric distribution before any threshold estimation methods are applied, so the corresponding method for transforming data is required. In this study, a new method for transformation of stream sediment geochemical data is provided. Firstly, the samples are classified by k-means method into different clusters, samples in each of which are thought to be of the same background. Then samples in each cluster are centered logratio transformed. Finally, the data after processed are tested and they all satisfy normal distributions. Furthermore, a stream sediment geochemical data set of a district in Hunan, China is taken as an example. Maps of anomalies of raw and transformed metallogenic Pb, Zn, Cu and W are portrayed respectively for comparison. The results show that anomalies of raw data correspond worse with the known deposits. By contrast, the method of mapping anomalies with transformed data performs better.