Feature Selection for Urban Land-Cover Classification using Landsat-7 ETM+ Data

Abstract

We report here the results of a study carried out to reduce the dimensionality of Landsat-7 Enhanced Thematic Mapper Plus (ETM+) digital data by principal component analysis, and generating a band triplet with maximum optimum index factor (OIF) value for developing land-cover map over a metropolitan city through Gaussian maximum likelihood algorithm. The performance of the thematic maps, thus generated from these three data sets, was done by a systematic accuracy assessment. Results indicate that a band triplet (ETM+ band 2, 4 and 5) with the maximum optimum index factor (OIF) value, and an overall accuracy of 97.5% and a kappa accuracy value of 0.9656 outperformed other two datasets viz. original 6-reflective bands of ETM+ data and a PC triplet (PC1, PC2 and PC3). The overall and a kappa accuracies values for original 6-reflective bands of ETM+ data have estimated as 96.7% and 0.9541, respectively. For a PC triplet (PC1, PC2 and PC3) these values are 94.17% and 0.9210, respectively indicating thereby the potential of transformed data in generating improved land cover information of an urban environment. The methodology and the results are discussed in detail.