Abstract

Analysis of Clustering and Classification Techniques for Mining of Remotely Sensed Imagery

Tasneem Akhtar
National University of Sciences & Technology, Institute of Information Technology,
Pakistan
tasneem.akhtar@niit.edu.pk

Muhammad Umer
National University of Sciences & Technology, Institute of Information Technology,
Pakistan
m.umer@niit.edu.pk

Rizwan Bulbul
National University of Sciences & Technology, Institute of Information Technology,
Pakistan
rizwan.bulbul@nu.edu.pk

Tashfeen Khan
National University of Sciences & Technology, Institute of Information Technology,
Pakistan
tashfeen@niit.edu.pk

Rizwan Ahmad
National University of Sciences & Technology, Institute of Information Technology,
Pakistan
rahmad@niit.edu.pk


Abstract :
The satellite sensors captures large amount of data in the form of Remotely Sensed Imagery (RSI) and continuously transmit it to ground stations. Enhancements in remote sensing technology, coupled with a parallel decrease in cost of data storage has resulted in rapidly mounting RSI data sets. Trend analysis, similarly search and other decision support analysis tasks have motivated researchers from several domains to work on RSI. In this backdrop, application of data mining - the suite of techniques applied to detect patterns in large data sets – is only natural. Over the last two decades, data mining techniques that have gained commercial acceptance mainly fall into clustering and classification domains. In this paper experiences are reported with the application of popular clustering and classification techniques for similarity searching in RSI. The primary goal of this paper is to investigate the effectiveness of these techniques in RSI domain on the basis of well-known quality indices. The clustering techniques analyzed in this study are K-means, ISODATA and agglomerative hierarchal clustering. Classification techniques include Mahlanobis Distance Classification, Minimum Distance Classification and NDVI classification. We apply these techniques on a test-bed of several hundred satellite images. Clustering techniques are analyzed on the basis of Separation, Cohesion (Sum of Squares Error) and Co-phenatic Coefficient indices, while classification techniques have been analyzed on the basis of precision and recall. Our analysis show several interesting results and provide new insights into the effectiveness of well-known data mining techniques in RSI mining domain.