Change Detection Using Unsupervised Learning Algorithms for Delhi, India

The effectiveness of the three types of unsupervised learning techniques for change detection in water, vegetation and built-up land cover classes of a part of Delhi region in India has been analyzed. A total of eight images of Landsat TM and ETM+ from year 1998 to 2011 were preprocessed for atmospheric corrections. Subsequently three features, Soil Adjusted Vegetation Index (SAVI), Modified Normalized Difference Water Index (MNDWI), and Builtup from Normalized Difference Built-up Index (NDBI) were extracted at the preprocessing stage. The three clustering algorithms kmeans, fuzzy c mean and expectationmaximization were selected to represent the partition based, fuzzy, and probability based technique respectively. The three algorithms were implemented to cluster the pixels of all the eight images using the features SAVI, MNDWI and NDBI. The Silhouette coefficient was used to evaluate the cluster quality that takes into consideration both intra-cluster and inter-cluster distance between clusters. The outcome of clustering has been quantified in terms of the percentage of total pixels grouped in each of the three clusters indicating vegetation, urban and water. Change detection has been performed comparing the outcomes of clustering done on each of the eight images.