Hyperspectral Satellite Image analysis: Image Processing and Classification using extracted pure pix
The study was carried out for Indian capital city Delhi using Hyperion sensor onboard EO-1 satellite of NASA. After MODTRAN-4 based atmospheric correction, MNF, PPI and n-D visualizer were applied and endmembers of 11 LCLU classes were derived which were employed in classification of LULC. To incur better classification accuracy, a comparative study was also carried out to evaluate the potential of three classifier algorithms namely Random Forest (RF), Support Vector Machines (SVM) and Spectral Angle Mapper (SAM). The results of this study reemphasize the utility of satellite borne hyperspectral data to extract endmembers and also to delineate the potential of random forest as expert classifier to assess land cover with higher classification accuracy that outperformed the SVM by 19% and SAM by 27% in overall accuracy. This research work contributes positively to the issue of land cover classification through exploration of hyperspectral endmembers. The comparison of classification algorithms’ performance is valuable for decision makers to choose better classifier for more accurate information extraction.
Keywords: Endmembers extraction, Hyperspectral Image classification, Random Forest - Ensemble classifier, SVM, SAM.
The study evaluated the applicability of satellite hyperspectral remote sensing imagery for hyperspectral endmembers extraction as well as possibilities of better land cover classification through machine learning and spectra angle based classifiers. Two machine learning techniques, support vector machines (SVM) and random forest (RF) with spectral angle mapper (SAM) were appraised and their suitability for better LCLU classification was determined. The following conclusions were drawn:
Due to its spectral resolution, the satellite borne hyperspectral data have the capability to provide spectrally pure pixels (endmembers) which are comparable to spectroradiometer spectra and available spectral libraries. These endmembers were successfully applied to derive subclasses of crops, water and human settlements. Accurate and informative LCLU maps were generated through machine learning (Random Forest) techniques with the highest overall accuracy of 88% (Kappa = 0.86).
The minimum noise fraction (MNF) transform is a very useful technique for hyperspectral data dimensionality reduction, noise removal, and spectral smile identification. Due to its data dimensionality reduction characteristic, computation time also reduces.
The study strengthens the fact that promising machine learning algorithms are indispensable tools for better classification and mapping. Machine learning classifiers especially Random Forest proved to be more accurate and outperformed both SVM and SAM classifiers. This type of classifier algorithms comparison may be useful for natural resource managers and environmentalists who deal with classified land covers categories to resolve the concerned issues because better classification accuracy will provide better information.