GAN vs. Traditional Methods: A Multi-Scale Performance Evaluation in Satellite Image Classification

Effective data augmentation is essential for enhancing machine learning model performance in satellite image classification. This study presents a multi-scale evaluation of Generative Adversarial Networks (GANs) compared to traditional augmentation methods, examining their effectiveness across high, medium, and low-resolution satellite imagery using the WHU-RS19, RSD46-WHU, and EuroSAT datasets. Our findings reveal that GANs significantly boost classification accuracy, with improvements of up to 8.23% in high-resolution images, 6.49% in medium-resolution images, and a modest 0.61% in low-resolution scenarios compared to traditional methods. Traditional techniques maintain competitive performance at lower resolutions, where simpler operations and reduced computational demands are advantageous. This research highlights the resolution-dependent benefits and constraints of each method, offering practical guidelines for selecting optimal augmentation strategies to enhance accuracy and generalizability in remote sensing applications.