Comparison of Three Distances In K-Means Clustering On Satellite Imagery

One of the preliminary works in the field of flood prediction, due to heavy rainfall, is the detection and identification of convective clouds using satellite imagery. Thermal infra-red (TIR) band images have been extensively used for this purpose. In order to identify the convective cloud, the image has to be clustered so that cloudy pixels can be identified. In this paper k-means clustering has been used for clustering pixels in a TIR image. From the image, four features such as mean, standard deviation, entropy, and busy-ness were obtained. Based on these features, clouds were segmented using k-means clustering algorithm. Finally, using a threshold value, cloudy pixels are extracted. Generally Euclidean distance is used in k-means clustering, but in this paper two more types of distances, Manhattan and Mahalanobis, have been used and the results have been observed using skill score analysis.