Hyperspectral imaging sensors capture data across hundreds of contiguous narrow spectral bands, allowing precise material identification that standard RGB cameras cannot achieve. However, this high dimensionality introduces the Hughes phenomenon, where classification accuracy degrades unless the number of training samples increases exponentially with the dimensionality (number of bands). As a result, collecting ground truth for labelled datasets becomes extremely expensive, and therefore, reducing data dimensionality by means of wavelength selection became mandatory. The proposed thesis presents a two-phase study of hyperspectral image (HSI) classification on the widely-used Salinas and Indian Pines benchmark datasets. In the first phase, the CCARS (Competitive Calibration Adaptive Reweighted Sampling) framework is extended to support non-linear classifiers, specifically Support Vector Machine with Radial Basis Function kernel (SVM-RBF) and Random Forest (RF).