This work presents a new methodology for the unsupervised classification of remotely sensed images. Differently of the traditional methods of unsupervised classification, such as K-means and ISODATA, which use only partitional clustering techniques, the proposed methodology accomplishes the automatic classification of images through an innovative approach applying the Kohonen Self-Organizing Map (SOM) together with an agglomerative hierarchical clustering method. The key point of the proposed method is to execute the clustering process through a set of prototypes of the SOM instead of analyzing directly the original patterns of the image. This approach significantly reduces the complexity of the analysis becoming possible the use of techniques that normally are considered impracticable for the digital processing of remotely sensed images, such as hierarchical clustering methods and cluster validity indexes. Using the SOM, the proposed method maps the original patterns of the image to a set of neurons arranged in a two-dimensional lattice searching to preserve the probability distribution and the topology of the input space. Subsequently, an agglomerative hierarchical clustering method with restricted connectivity is applied on the lattice of neurons previously trained, generating a simplified dendrogram for the image data. Each level of the dendrogram corresponds to a different configuration of clusters of neurons (or prototypes) of the SOM that can be used to represent the classes on which the original image will be classified. Applying modified versions of cluster validity indexes the method automatically determines the ideal number of clusters of the image not demanding that the user previously defines the quantity of classes to execute the classification process. The experimental results show an application example of the proposed method on a test image and its performance is compared with the K-means algorithm.