SVM Kernel Based Method for Satellite Image Classification

Satellite image classification plays a vital role in environmental monitoring, land cover mapping, and resource management. However, the high dimensionality and nonlinear patterns in such data demand robust machine learning approaches. This study investigates the effectiveness of Support Vector Machines (SVMs) with different kernel functions for multiclass satellite image classification, targeting four categories: Cloudy, Desert, Green Area, and Water. The research evaluates four kernel types—Linear, Polynomial, Radial Basis Function (RBF), and Sigmoid—combined with five multiclass strategies: One-Against-One (OAO), One-Against-All (OAA), Directed Acyclic Graph SVM (DAGSVM), Weston-Watkins (WW) SVM, and Crammer-Singer (CS) SVM. A dataset of 5,631 images is preprocessed using Histogram of Oriented Gradients (HOG) for feature extraction, followed by normalization and stratified splitting. Rigorous hyperparameter tuning is conducted across 195 experiments, with performance assessed via the Kappa coefficient and F-measure to ensure statistical reliability. The study emphasizes the interplay between kernel selection and multiclass strategies, addressing gaps in optimizing SVM frameworks for remote sensing applications. By systematically comparing these approaches, the research aims to identify configurations that balance accuracy, computational efficiency, and adaptability to non-linear data structures. Limitations include the dataset’s geographic scope, suggesting future work with diverse spatial-temporal data or hybrid models integrating deep learning. The findings are expected to guide practitioners in deploying SVMs for scalable and precise land cover classification.