Journal of Pattern Recognition Research

This paper presents a technique for performing unsupervised segmentation of satellite images using a `sampling-resampling' based Bayesian learning method. The multi-band pixel values of the satellite image are grouped into clusters that are modeled using Gaussians. The parameters of this Gaussian mixture model are learnt using a Bayesian approach. It is well known that Bayesian estimates of model parameters are often more reliable and regularized, since Bayesian learning yields a complete posterior distribution of model parameters, and not just singleton estimates that are obtained using frequentist techniques like Maximum Likelihood. This paper has the novelty of using a `sampling-resampling' implementation of Bayesian learning, which is easier to implement as compared to numerical integration techniques that are conventionally used. The proposed method is unsupervised in the sense that no separate training images are used to learn the model parameters. It can be observed in our results that the clustering/segmentation results obtained using the proposed technique have a better correspondence to actual land features in the satellite images, as compared to results of established clustering techniques like the K-means algorithm.