Hierarchical clustering based on the finite mixture model（FM） has shown very good performance in a number of fields.However

it generally requires storage and computing at least proportional to the square of the dimension of observations

so that its application to large datasets has been hindered by time and memory complexity.Another

multispectral images provide detailed data with information in both the spatial and spectral domains.But many clustering methods for multispectral images are based on a per-pixel classification

while uses only spectral information and ignores spatial information.In this work

a new hierarchical clustering based on GFM model

suitable for large datasets

e.g.

multispectral remote sensing images

is proposed.This algorithm integrates with GFM model with Markov random field.The number of clusters is automatically identified by using the pseudolikelihood information criterion（PLIC）.An oversegmented image is obtained by a simple K-means clustering method.Instead of starting with singleton clusters

hierarchical clustering is applied on the oversegmented image.Initial parameters of component densities of GFM model can be easily extracted.At last

the accuracy of the algorithm is quantitatively evaluated through simulated test image generated by using Gibbs sampler.The experiment show a superior performance compared to several other methods

such as K-means and classical hierarchical clustering based on the classical FM model.Its validity is also illustrated by using a polarimetric SAR image of Flevoland in the Netherlands.