Inverse Domain Adaptation for Remote Sensing Images Using Wasserstein Distance

Abstract

In this work, an inverse domain adaptation (IDA) method is proposed to cope with the distributional mismatch between the training images in the source domain and the test images in the target domain in remote sensing. More specifically, a cycleGAN structure using the Wasserstein distance is developed to learn the distribution of the remote sensing images in the source domain before the images in the target domain are transformed into similar distribution while preserving the image details and semantic consistency of the target images via style transfer. Extensive experiments using the GF1 data are performed to confirm the effectiveness of the proposed IDA method.