Optimal Parameter Selection for High Accuracy Area Classification in Hyper Spectral and Landsat-8 Imagery with Seasonal Change Detection

Abstract

The thesis proposes area classification after construction of datasets with pre-processing for Hyperion Hyperspectral, Operational Land Imager (OLI) and Advanced Land Imager (ALI) orthoimages. The techniques perform comparative analysis of Hyperion Hyperspectral, OLI and ALI orthoimages in terms of high Signal to Noise (SNR), spectral band configuration, technical superiority, improved system design and high radiometric resolution. The thesis further proposes criteria for selection of parameters like gamma parameter, penalty parameter, pyramid parameter and classification probability threshold to achieve higher classification accuracies of Hyperion Hyperspectral, OLI and ALI Satellite orthoimages by using Support Vector Machine (SVM), Spectral Angle Mapper (SAM) and Spectral Information Divergence (SID) classifiers. After performing the comparison, the thesis selects SVM as the most appropriate classifier in terms of overall accuracy, individual classes and the best orthoimagery i.e. Hyperion, OLI and ALI respectively. The thesis also presents application of classification accuracy assessment on Hyperion Hyperspectral, OLI and ALI orthoimages by using different classifiers i.e. SVM, SAM and SID. The thesis also proposes high accuracy based seasonal change detection analysis technique on Hyperion Hyperspectral, ALI and different datasets of OLI by using change detection matrix and difference maps. As a result of these contributions, 1x Journal and 6 x Conference papers duly peer reviewed have been published. The pre-processing of 242 bands of hyperspectral data results in 136 calibrated bands. Quick Atmospheric Correction (QUAC) are applied to Hyperion Hyperspectral and Fast Line-of-sight Atmospheric Analysis of Hyper cubes (FLAASH) are applied to OLI and ALI imagery respectively for atmospheric correction. Principal Component Analysis (PCA) is used for dimensional reduction of the hyperspectral data. PCA reveals that 99.94% of the hyperspectral data are contained in the first 15 Principal Components (PCs). Distinct spectral profiles are identified for all classes which are highly beneficial for feature identification and classification of images. Novel parameters are selected for high accuracy area classification in hyperspectral, OLI and ALI imagery via SVM, SAM and SID classification techniques. High accuracy based post classification change detection analysis is used on Hyperion Hyperspectral, ALI and different OLI datasets to produce difference maps which provide information not only about change of category but also type of change i.e. “from-to” of category of classes. Change detection matrix is also used which shows an overall decrease and increase of corresponding spatial extension of classes whereas diagonal elements of the change detection matrix show the unchanged pixels for the individual classes. The post classification technique is selected because of its ability for accurate change detection analysis of imagery of different sensors and its advantages over pre-classification methods that it compensates for variation in atmospheric correction and in conditions where the change is limited due to small rate of change. The results show that Hyperion hyperspectral and Landsat-8 OLI data achieved higher accuracies in mapping applications and high accuracy based post classification seasonal change detection analysis on different datasets on OLI results extraction of accurate change detection information as compared to previous Landsat satellite series.