Abstract

This study proposes the use of generative models (GANs) for augmenting theEuroSAT dataset for the Land Use and Land Cover (LULC) Classification task. Weused DCGAN and WGAN-GP to generate images for each class in the dataset. Wethen explored the effect of augmenting the original dataset by about 10% ineach case on model performance. The choice of GAN architecture seems to have noapparent effect on the model performance. However, a combination of geometricaugmentation and GAN-generated images improved baseline results. Our studyshows that GANs augmentation can improve the generalizability of deepclassification models on satellite images.

 

Quick Read (beta)