OPTIMUM BAND SELECTION FOR DISCRIMINATION OF VARIOUS LAND COVER FEATURES USING HYPERSPECTRAL DATA

Abstract

Hyperspectral sensors can produce data of fine spectral resolution for the identification of materials from satellite observations. As the spectral resolution increases, the capability to detect more detailed classes may also increase. The high dimensionality of Hyperspectral data is a known phenomenon that can cause imprecise class estimation if too many spectral bands are simultaneously used in the classification process. Dimensionality reduction may, therefore, be necessary to reduce the size and redundancy in remote sensing data, which in turn reduces the cost and complexity of image processing. The dissertation presented here is about the selection of waveband regions which best discriminate a number of land cover features. To perform dimensionality reduction four statistical techniques, namely Stepwise Discriminant Analysis (SDA), Principal Component Analysis (PCA), Mann Whitney U test and BBR2M, are used in this study. Each technique has specific criteria to select the optimal number bands, which have high seperability among various land cover classes. The dataset selected by each method shall be using the classification process to produce the classified map. SAM classifier is used in this study. Based on the accuracy of these classes effectiveness of each method shall be highlighted. Some bands are also selected by combining the bands selected by each method. Classification process is also applied on this dataset to determine its effectiveness. The study has been conducted on Hyperion dataset. The classification results indicate that bands selected from combination of all the four methods have highest accuracy i.e. higher discrimination capability. PCA also have good classification accuracy but it has more number of bands in comparison to other methods. Classification accuracy of all the methods is comparable to the classification accuracy of the dataset with all bands. This indicates that the band selection by using SDA, PCA, U Test and BBR2M have the high capability to discriminate various land covers.