EVALUATION OF SPECTRAL DISTANCE FOR EFFECTIVE FEATURE MAPPING IN A MULTICLASS PROBLEM

Abstract

Mapping of land surface features through satellite images is one important aspect of Remote sensing applications. The accuracy of this process depends on how the feature selection is carried out. In a multispectral satellite image, the separability of the features depends on the band combinations used. The Landsat image was considered in this study and the Histogram minimum method of atmospheric correction which is suitable for cloud free images. The image autocorrelation reveals that the Moran’s I obtained for the three LULC images are all positive values as 0.6368, 0.6167 and 0.8073 for the year 1986, 2000 and 2019 respectively which indicates tendency towards clustering. Using sample of training areas, Jeffries-Matusita distance was computed as a measure to check the degree of spectral separability on the bands used. This is to investigate if any pairs of land cover (LC) class consistently show similar spectral signature. This study presents a spectral separability analysis for LC features using Landsat-8 OLI/TIRS of 2019, Landsat 5 Thematic Mapper (TM) of 1986 and Enhanced Thematic Mapper (ETM) of the year 2000. This measurement reveals spectral overlapping between selected LC features and for the discrimination, statistical comparisons such as calculating the means of digital numbers (DNs), standard deviation of single band reflectance values and plotting the standard errors on all classes used. LTM 1986 shows pretty good separation in bands 3, 4, 5 & 7 while 1, 2 & 6 did not. Bands 3, 4, 5 & 7 in ETM 2000 also indicated good separability while 1, 2 & 6 clearly show poor spectral separability. Finally, L8 OLI returned bands 5, 6 & 7 with appreciable separation while bands 1, 2, 3, 4, & 8 did show good separation. These statistical results were proven to be good indicators of spectral similarity because the heterogeneity measure makes it more effective and objective