Cross-Sensor Adversarial Domain Adaptation of Landsat-8 and Proba-V\n images for Cloud Detection

Abstract

The number of Earth observation satellites carrying optical sensors with\nsimilar characteristics is constantly growing. Despite their similarities and\nthe potential synergies among them, derived satellite products are often\ndeveloped for each sensor independently. Differences in retrieved radiances\nlead to significant drops in accuracy, which hampers knowledge and information\nsharing across sensors. This is particularly harmful for machine learning\nalgorithms, since gathering new ground truth data to train models for each\nsensor is costly and requires experienced manpower. In this work, we propose a\ndomain adaptation transformation to reduce the statistical differences between\nimages of two satellite sensors in order to boost the performance of transfer\nlearning models. The proposed methodology is based on the Cycle Consistent\nGenerative Adversarial Domain Adaptation (CyCADA) framework that trains the\ntransformation model in an unpaired manner. In particular, Landsat-8 and\nProba-V satellites, which present different but compatible spatio-spectral\ncharacteristics, are used to illustrate the method. The obtained transformation\nsignificantly reduces differences between the image datasets while preserving\nthe spatial and spectral information of adapted images, which is hence useful\nfor any general purpose cross-sensor application. In addition, the training of\nthe proposed adversarial domain adaptation model can be modified to improve the\nperformance in a specific remote sensing application, such as cloud detection,\nby including a dedicated term in the cost function. Results show that, when the\nproposed transformation is applied, cloud detection models trained in Landsat-8\ndata increase cloud detection accuracy in Proba-V.\n