ChangeNet: Learning to detect changes in satellite images

Abstract

Change detection in temporal sequences of satellite images is an important component of many remote sensing applications such as land cover monitoring, urban expansion evaluation, forest degradation assessment, and mine site monitoring. The objective of this paper is to localize and identify relevant pixelwise changes in time-varying images taken at the same location. Detecting relevant change in images is difficult due to "unimportant" or "nuisance" forms of change such as illumination variation, shadows, occlusion, and possible seasonal changes. Traditional methods for change detection require sophisticated image preprocessing and possibly manual interaction. In this work, we present an end-to-end approach for dense change detection in satellite images by employing conditional Generative Adversarial Networks. We use the conditional GAN network to improve classification results by closing the gap between expected and predicted label distributions. Experimental results show that the proposed method achieves better performance compared with existing methods.