Initialization methods for remote sensing image clustering using K-means algorithm

Unsupervised remote sensing image classification algorithms are very sensitive to the initial conditions.Using the K-means algorithm as an example,the influence of five initialization methods on unsupervised classification algorithms is respectively compared by means of various experiments in remote sensing images.Although K-means is known for its robustness,it is widely reported in the literature that its performance depends upon initial clustering.A series of experiments are conducted to evaluate the performance of different initialization methods in terms of overall accuracy,Kappa coefficient,initial time and iteration number of convergence.The results of the experiments illustrate that the Kaufman initialization method outperforms the rest of the compared methods as they make the K-means more effective and more independent on initial clustering and suggest that the initial time of the Kaufman method can be reduced while maintaining the well results.The convergence speed of the K-means algorithm is also compared using each of the five initialization methods.In addition,the sensitivity of initialization methods in relation to the number of sampling and the image's size is analyzed.