GAN-Based Satellite Image Classification

Overview

A research project investigating how different GAN architectures can augment the EuroSAT satellite imagery dataset to improve land-use classification accuracy. Published at IEEE ICESC 2024.

My Role

Lead researcher — designed experiments, implemented GAN variants, trained classification models, and authored the paper.

Approach

• Implemented four GAN variants: InfoGAN, CGAN, CDCGAN, and ACGAN for synthetic satellite image generation.
• Augmented the EuroSAT dataset by 50% with generated samples across all land-use categories.
• Trained CNN classifiers on original vs. augmented datasets to measure classification improvement.
• Conducted rigorous evaluation using accuracy, precision, recall, and FID scores across all variants.

Outcome

CDCGAN achieved the highest classification accuracy; ACGAN excelled in precision and recall. Published at IEEE ICESC 2024. Cited by 4.