Abstract

Satellite images based on hyperspectral imagery (HSI) are required to undergo additional processing in order to be utilized in various applications. The classification of hyper-spectral images is a crucial and necessary process for mapping coordinates from the image coordinates. In this procedure, the ground control points (GCPs) need to be manually retrieved from the remotely sensed images, relying on the ground truth values. This particular step is known to be time-consuming. Therefore, this study introduces a method called super pixel based principal component analysis (Super-PCA) for the classification of multi-temporal hyperspectral satellite data. The suggested method aims to automatically extract ground control points (GCPs) and reduce computational time, thereby enhancing classification accuracy. Principal Component Analysis (PCA) is a commonly employed methodology in satellite data analysis for the purpose of feature extraction. However, it is worth noting that the process of extracting features using PCA can be computationally intensive, resulting in slower performance. In order to address the limitations of Principal Component Analysis (PCA), a novel approach known as Super-PCA has been suggested and implemented. However, it is important to note that this technique has mostly been used to images with a relatively low number of features and has not yet been extended to the domain of satellite imagery. In this research, the Super-PCA technique has been enhanced in terms of phase angle for the purpose of feature extraction in multi-temporal hyperspectral satellite imagery at six different levels. The Support Vector Machine (SVM) algorithm is commonly employed in the context of non-linear multi-class classification tasks. The efficacy of Support Vector Machines (SVM) is contingent upon the careful selection of an appropriate kernel. Therefore, the Fuzzy Support Vector Machine (F-RVM) is suggested as a method for selecting the appropriate kernel in satellite imagery analysis, taking into account the resolution and intensity of the features. The obtained results are compared with a range of conventional procedures, demonstrating superior performance.

Description