Wave propagation laws are highly linked with environmental nature (city, country, mountains, etc...). Within the framework of a cell net planning in radiocommunication, we are interested in determining classes, homogeneous enough, upon which specific prediction models of radio electrical field can be applied. Various algorithms for unsupervised vector quantization exist and do not yield exactly the same result on the same problem because quantization can be done from different points of view. To better understand this phenomenon, this article presents evaluation of unsupervised neural networks, among the most useful for quantization, applied to a real-world problem. A particular interest is given to techniques that improve data analysis. The use of Mahalanobis' distance allows an assignment independently of the data correlation. The study of class dispersion and homogeneity using data structure and statistical analysis put in a prominent the global properties of each algorithm. Finally, we discuss the interest of these methods on a real problem of clustering linked to radiocommunication.