[Paper Review] Image Augmentation for Satellite Images

Oluwadara Adedeji, Peter Owoade+2 more|arXiv (Cornell University)|Jul 29, 2022

Remote-Sensing Image Classification6 citations

TL;DR

This study proposes using GANs—specifically DCGAN and WGAN-GP—to generate synthetic satellite images for data augmentation in Land Use and Land Cover (LULC) classification. It finds that GAN-generated images improve model generalizability without degrading performance, and combining them with geometric augmentation yields better results than either method alone.

ABSTRACT

This study proposes the use of generative models (GANs) for augmenting the EuroSAT dataset for the Land Use and Land Cover (LULC) Classification task. We used DCGAN and WGAN-GP to generate images for each class in the dataset. We then explored the effect of augmenting the original dataset by about 10% in each case on model performance. The choice of GAN architecture seems to have no apparent effect on the model performance. However, a combination of geometric augmentation and GAN-generated images improved baseline results. Our study shows that GANs augmentation can improve the generalizability of deep classification models on satellite images.

Motivation & Objective

• To address the challenge of limited training data in satellite image classification, particularly for the EuroSAT dataset.
• To investigate whether generative models like GANs can produce realistic synthetic satellite images that improve model robustness.
• To evaluate the impact of GAN-generated images on model performance compared to traditional geometric augmentation.
• To determine whether different GAN architectures (DCGAN vs. WGAN-GP) yield measurable differences in classification accuracy.
• To explore the potential of combining GAN augmentation with geometric augmentation for improved generalization in low-data regimes.

Proposed method

• Trained DCGAN and WGAN-GP on the EuroSAT dataset to generate synthetic satellite images for each of the 10 land use classes.
• Applied a 10% data augmentation rate by adding GAN-generated images to the original training set for each class.
• Combined GAN-generated images with standard geometric augmentation techniques (e.g., flipping, rotation, zooming) to create hybrid augmented datasets.
• Trained VGG16 and ResNet50 models on both original and augmented datasets using cross-entropy loss and Adam optimization.
• Used early stopping and learning rate scheduling to stabilize training, with validation accuracy monitored over 25 training epochs.
• Evaluated model performance using accuracy on the test set, comparing baseline, GAN-augmented, and geometric-augmented configurations.

Experimental results

Research questions

• RQ1 Does augmenting the EuroSAT dataset with GAN-generated satellite images improve classification accuracy for LULC tasks?
• RQ2 Is there a significant difference in model performance when using DCGAN versus WGAN-GP for data augmentation?
• RQ3 How does combining GAN-based augmentation with geometric augmentation affect model generalizability compared to using either method alone?
• RQ4 Does GAN-based augmentation help reduce overfitting in deep learning models trained on small satellite image datasets?
• RQ5 Can GAN-generated images serve as a viable alternative or complement to traditional data augmentation in remote sensing applications?

Key findings

• GAN-augmented datasets achieved comparable or slightly improved test accuracy compared to the original dataset, with no degradation in performance.
• The choice between DCGAN and WGAN-GP had no significant impact on model accuracy, suggesting both generate images of sufficient quality for the task.
• A combination of geometric augmentation and GAN-generated images yielded better performance than geometric augmentation alone, indicating synergistic benefits.
• The improvement from GAN augmentation was most pronounced in scenarios with limited data, as the method helps reduce overfitting and improves model generalization.
• Even with high baseline accuracy (e.g., 99.04% for ResNet50), GAN augmentation did not degrade performance, confirming the quality of generated images.
• The results suggest that GAN-based data augmentation is a viable and effective strategy for improving deep learning models in low-data satellite image classification tasks.