Land-Cover Classification of WorldView-3 Multispectral Imagery with K-Means, GMM, and SOM after Dimensionality Reduction

Abstract

Multispectral satellite imagery provides rich information about the Earth's surface. However, strong correlations among spectral bands often introduce redundancy and increase computational complexity. In this study, high-resolution WorldView-3 imagery is used to analyze spectral band relationships and classify different types of land cover. Principal Component Analysis (PCA) is applied to reduce the original eight spectral bands to three orthogonal components while preserving over 95% of the total variance. The reduced data is then used as input to three classification algorithms, K-Means, Gaussian Mixture Model (GMM), and Self-Organizing Map (SOM), to group similar land types. The results show that classification with dimensionality reduction is twice as fast compared to using all spectral bands. K-Means performs best with a Silhouette score of 0.4, a Davies-Bouldin index of 0.8, and a Calinski-Harabasz score of 103570. GMM creates smoother but overlapping clusters, while SOM captures general patterns but the groups are more spread out. Finally, K-Means and SOM successfully separate visually similar surfaces, such as water and shaded vegetation, which are difficult to distinguish through visual inspection, whereas GMM achieves comparable separation after combining it with the other methods. Incorporating NDVI threshold further improves the classification.