Hyperspectral and Multispectral Remote Sensing Data Fusion for Classification of Complex-Mixed Land Features Using SVM

In the present paper, classification and analysis of complex land features on the combined outcome of the high spectral resolution Hyperion-Hyperspectral Remote Sensing (HRS) data and high spatial resolution Resourcesat-II Linear Imaging Self-Scanning System IV (LISS-IV) multispectral data were investigated. The traditional way of satellite image fusion is based on high spatial resolution panchromatic image and low spatial-spectral resolution multispectral image. However, in the current study, a novel approach via considering HRS and LISS-IV multispectral data is proposed for classification of complex features of earth surface. The used multi-date, multi-sensor and multi-resolution satellite imagery was acquired on \(20^{th}\) March, 2015 and \(28^{th}\) February, 2014 of HRS and LISS-IV data having spatial resolution 30 m and 5.8 m respectively. Three pixel level image fusion algorithms were computed such as Gram-Schmidt Transform (GST), Principal Component Spectral Sharpening Transform (PCSST) and Color Normalized Spectral Sharpening (CNSS) for fusion of datasets. The quality of the fusion algorithms has been estimated on the classification accuracy of mixed features. Moreover, the performance of three fusion algorithms was compared with the classification results. The assessment results of fused data using all the methods were acceptable in view of spatial-spectral accuracy of data. The Support Vector Machine (SVM) approach with its Gaussian Radial Basis Function (GRBF) kernel was implemented for classification of original and fused data. In conclusion, the SVM algorithm resulted accurate with 97.65, 97.47, 96.30, 86.20 and 74.44% accuracy for CNSS fused, PCSST fused, GST fused, and original multispectral and original hyperspectral data respectively and proved very robust method for mixed feature classification.