Nowadays, geographic information becomes too complex and abundant especially due to the development of remote sensing devices with high frequent acquisition rates, high spectral and spatial resolutions, etc. To make this flow of data manageable and exploitable, we need to represent it in a way it could be understood by humans and machines. Ontologies are considered as a valuable support for knowledge representation but they need to be built first. Building ontologies can be done either manually or automatically. Building ontologies manually is expensive and tiresome since we need experts. Moreover, it is intractable. Automatic ontology building is more suited to handle intractability, and can be reduced as an enrichment process, i.e. enriching an existing core ontology modelling ontological information by new concepts coming from other geographic ontologies. Alignment of concepts coming from two distinct ontologies is a key issue in the enrichment process and deeply affects the quality of the resulting ontology. The alignment of two ontologies aims at putting in correspondence concepts of a main ontology with concepts of a target ontology using similarity measures, as well as a set of parameters such as weights or thresholds, and a set of external resources such as thesauri. In the literature, there are many models and methods for ontology alignment. They mainly differ with respect to the similarity measures they use as well as the way these measures are combined. Most of the alignment methods do not deal with the problem of the correlation existing between the similarity measures they use, nor to the fact that the degree to which each measure contributes to the overall similarity (i.e. their weights) can vary depending on the ontologies involved in the alignment process. In this chapter, we address the first critical issue allowing to better deciding which similarity measures should we consider to assess the true similarity between concepts. Our proposal consists of using feature selection methods, in order to select a reduced set of relevant similarity measures.