Series Info:Proceedings of the European Conference on Data Mining 2014 and International Conferences on Intelligent Systems and Agents 2014 and Theory and Practice in Modern Computing 2014 - Part of the Multi Conference on Computer Science and Information Systems, MCCSIS 2014
Type:Conference Paper
Keywords:October University for Modern Sciences and Arts
,
University for Modern Sciences and Arts
,
MSA University
,
جامعة أكتوبر للعلوم الحديثة والآداب
,
Clustering
,
Hyperspectral images
,
Image processing
,
K-means
,
Remote sensing
,
Computation theory
,
Data mining
,
Image processing
,
Independent component analysis
,
Intelligent agents
,
Intelligent systems
,
Reflection
,
Remote sensing
,
Spectroscopy
,
Statistics
,
Analysis accuracy
,
Clustering
,
Hyper-spectral images
,
K-means
,
Random kernels
,
Reflectance values
,
Spatial resolution
,
Spectral feature
,
Pixels
Abstract:
The attention given recently for Hyperspectral Images (HSI) in remote sensing was due to its spectral nature used in earth surface exploration. Recently, researches proposed approaches that combine features of spatial nature with spectral nature to enhance HSI analysis accuracy. The reason behind this orientation is that contiguous pixels mostly share spectral features due to the low spatial resolution. In this paper, authors follow this orientation in clustering HSI using K-means. The proposed approach uses kernels, group of spatially neighbored pixels, to build profile for each cluster. The profile preserves cluster's spectral nature through its bands' Box-Plots that were extracted from selected kernels. The approach starts by selecting K kernels from the given image scene randomly instead of generating random kernels. Then a profile for each cluster is built using the selected kernel pixels' spectrums. The profile consists of b box-plots, where b is the number of bands. Each band box-plot interprets the spread of contiguous pixels' reflectance values in that band. Each pixel in the image will join the nearest cluster. The distance to be measured between any given pixel spectrum and cluster's centroid is replaced by counting how many reflectance values have been considered outlier to their corresponding band's box-plot in that cluster profile. The pixel will join the cluster with minimum outlier count. The profiles are updated iteratively using the new pixels distribution. Data sets used in the experiments are captured by Hyperion Earth Observer 1 (EO-1) sensor. Copyright � 2014 IADIS Press All rights reserved.