Hyperspectral image (HSI) classification remains challenging due to high spectral dimensionality, redundancy among bands, and limited labeled samples, particularly in high–spatial-resolution agricultural and coastal environments. A comparative dimensionality-reduction and classification framework is presented and evaluated on two distinct hyperspectral scenarios: the WHU-Hi benchmark dataset acquired using UAV-borne hyperspectral sensors for precision crop classification, and a mangrove hyperspectral dataset collected over the Henry Island coastal ecosystem. The hyperspectral data cubes, consisting of hundreds of spectral bands and over 386,000 labeled samples, are transformed using Principal Component Analysis (PCA), a Modified PCA (MPCA) strategy with standardized variance normalization, and Kernel PCA to obtain compact and discriminative feature representations. The reduced feature sets, limited to 30 principal components, are evaluated using five supervised machine-learning classifiers, including Random Forest, Light Gradient Boosting Machine, Extreme Gradient Boosting, Support Vector Machine, and K-Nearest Neighbors. Experimental results indicate that PCA- and MPCA-based features achieve consistently high classification performance across all classifiers. The highest overall accuracy of 87.96% is obtained using SVM with PCA/MPCA features, while Random Forest and KNN achieve accuracies of 85.18% and 84.34%, respectively. Notably, MPCA achieves equivalent classification accuracy to conventional PCA while reducing feature extraction time by more than 60%, demonstrating superior computational efficiency. Overall, the framework provides an effective and computationally efficient solution for UAV-based crop classification and large-scale coastal ecosystem monitoring using hyperspectral imagery.
Chowdhury et al. (Wed,) studied this question.