Multiple K Means++ Clustering of Satellite Image Using Hadoop MapReduce and Spark

TL;DR

This paper compares multiple K-Means++ clustering algorithms on satellite images using Hadoop MapReduce and Spark, focusing on efficiency and validation of clustering results on big data platforms.

Contribution

It introduces a method to run multiple K-Means++ clustering algorithms simultaneously on big data platforms with validation, enhancing satellite image analysis.

Findings

Spark outperforms MapReduce in processing speed.

Simultaneous clustering reduces computation time.

Validation with Silhouette Index confirms clustering quality.

Abstract

Clustering of image is one of the important steps of mining satellite images. In our experiment we have simultaneously run multiple K-means algorithms with different initial centroids and values of k in the same iteration of MapReduce jobs. For initialization of initial centroids we have implemented Scalable K-Means++ MapReduce (MR) job [1]. We have also run a validation algorithm of Simplified Silhouette Index [2] for multiple clustering outputs, again in the same iteration of MR jobs. This paper explored the behavior of above mentioned clustering algorithms when run on big data platforms like MapReduce and Spark jobs. Spark has been chosen as it is popular for fast processing particularly where iterations are involved.