Enhancing Change Detection in Remote Sensing: Integrating Synthetic Data with Semi-Supervised Learning

Abstract

Change detection (CD) in remote sensing is a crucial yet challenging task, particularly due to the labor-intensive nature of labeling bi-temporal images. We introduce a novel framework that leverages synthetic datasets, style transfer, and semi-supervised learning to enhance CD model performance while reducing the dependency on labeled data. Our approach begins with a GAN-based style transfer model that transforms synthetic images to align with real-world scenarios, narrowing the domain gap. These transformed images, combined with a small amount of labeled real data, are used for supervised training to build a robust initial model. We then apply a mean teacher model to integrate unlabeled real images, allowing for effective semi-supervised learning. Our method achieves state-of-the-art performance on the LEVIR and WHU-CD datasets, demonstrating its robustness and accuracy across diverse geographical and temporal conditions.