In this article we propose an automatic procedure that performs the zonation of urban areas based on a set of horizontal-to-vertical (h/v) spectral ratios from ambient noise recordings collected at many measurement points. The technique searches for the connected areas where the similarity among the spectral ratios is maximized. The problem is posed as one of the optimal partitioning of the Delaunay triangulation of the available measurement points. The technique explores and tries to partition some random variations of a Euclidean minimum spanning tree of the triangulation. The optimization is performed using a genetic algorithm. The technique is applied to the zonation of the town of Vittorio Veneto in northeastern Italy and to a synthetic data set that tries to simulate the effects of some typical geological conditions.