Siamese Networks with Transfer Learning for Change Detection in Sentinel-2 Images

The Earth’s surface is constantly changing due to various anthropogenic and natural causes. Leveraging machine learning to monitor land cover changes over time may provide valuable information on the transformation of the Earth’s environment. This study focuses on the discovery of land cover changes in bi-temporal, Sentinel-2 images. In particular, we rely on a Siamese network trained with labelled, imagery data of the same Earth’s scene acquired with Sentinel-2 at different times. Subsequently, we adopt a transfer learning strategy to adapt the Siamese network to Sentinel-2 data acquired in any new unlabeled scene. To deal with the lack of change labels in the new scene, transfer learning is performed with change pseudo-labels estimated in the new scene in unsupervised manner. We assess the effectiveness of the proposed change detection method in two couples of images acquired with Sentinel-2, at different times, in the urban areas of Cupertino and Las Vegas.