Transfer Learning With Generative Adversarial Networks: Pretrained and Fine-Tuning Approaches in Remote Sensing Application

Abstract

The integration of transfer learning approaches with a GAN is a significant improvement in the state-of-art in the application of remote sensing especially given that the process is becoming compounded by the presence of large volume of the EO data. The chapter presents the correlation between transfer learning and GANs regarding the pretrained models and fine-tuning approaches to determine the potential for the transformation of remote sensing data analysis. The exploration of transfer learning with GANs in remote sensing focuses on two primary approaches: base models and the methods of transfer learning. Therefore, the methodologies of pretraining GANs revealed that much fewer training data and computational resources are needed for certain remote sensing applications. Interestingly, the fine-tuning approaches allows for the transfer learning of these released GANs on the distinctive attributes of the RS data improving its performance in several functionalities such as image categorization, object detection, and data amalgamation.