Analyzing the Impact of Geospatial Derivatives on Domain Adaptation with CycleGAN

Abstract

In scenarios when you have two sets of satellite images from distinct domains, CycleGAN provides image-to-image translation for automatic land-cover mapping. Style transfer and domain adaptation are two areas that heavily rely on the high-quality and varied picture transformation outcomes that deep learning can produce. CycleGAN's fundamental ideas are widely applied in many fields, including Geographic Information Systems. In order to improve CycleGAN's domain adaptation, we investigate the possibilities of using additional geospatial derivatives in this work along with multispectral bands. In terms of datasets, we explore the potential of using digital surface models and vegetation indices to enhance domain adaptation. This study also enhances CycleGAN's domain adaptation by comparing histogram equalization and K-means clustering as post-processing steps. Our results show that these modifications have the potential to yield better outcomes at land cover classification after domain adaptation. The proposed approach has been tested on Potsdam and Vaihingen datasets with similar class labels but variation in land-cover features. Further research will consolidate efforts to enhance automatic classification of land cover classes without any training.