GANs Based Conditional Aerial Images Generation for Imbalanced Learning

In this paper, we examine whether we can use Generative Adversarial Networks as an oversampling technique for a largely imbalanced remote sensing dataset containing solar panels, endeavoring a better generalization ability on another geographical location. To this cause, we first analyze the image data by using several clustering methods on latent feature information extracted by a fine-tuned VGG16 network. After that, we use the cluster assignments as auxiliary input for training the GANs. In our experiments we have used three types of GANs: (1) conditional vanilla GANs, (2) conditional Wasserstein GANs, and (3) conditional Self-Attention GANs. The synthetic data generated by each of these GANs is evaluated by both the Fréchet Inception Distance and a comparison of a VGG11-based classification model with and without adding the generated positive images to the original source set. We show that all models are able to generate realistic outputs as well as improving the target performance. Furthermore, using the clusters as a GAN input showed to give a more diversified feature representation, improving stability of learning and lowering the risk of mode collapse