During the past decade, the use of DNA for forensic applications has been extensively implemented for plant and animal species, as well as in humans. Tracing back the geographical origin of an individual usually requires genetic assignment analysis. These approaches are based on reference samples that are grouped into populations or other aggregates and intend to identify the most likely group of origin. Often this grouping does not have a biological but rather a historical or political justification, such as "country of origin". In this paper, we present a new nearest neighbour approach to individual assignment or classification within a given but potentially imperfect grouping of reference samples. This method, which is based on the genetic distance between individuals, functions better in many cases than commonly used methods. We demonstrate the operation of our assignment method using two data sets. One set is simulated for a large number of trees distributed in a 120km by 120km landscape with individual genotypes at 150 SNPs, and the other set comprises experimental data of 1221 individuals of the African tropical tree species Entandrophragma cylindricum (Sapelli) genotyped at 61 SNPs. Judging by the level of correct self-assignment, our approach outperformed the commonly used frequency and Bayesian approaches by 15% for the simulated data set and by 5-7% for the Sapelli data set. Our new approach is less sensitive to overlapping sources of genetic differentiation, such as genetic differences among closely-related species, phylogeographic lineages and isolation by distance, and thus operates better even for suboptimal grouping of individuals.