Change detection in remote sensing imagery typically requires large amounts of manually labeled image pairs, which is expensive and time-consuming. This thesis uses the BTC framework with a DINOv3 encoder as the basis and trains the model exclusively on synthetic changes, studying in detail a DRAEM-like approach with an organic mask shape and soft content blending. Results across three satellite imagery datasets show that simpler methods lead to shortcut learning, while the DRAEM-like approach considerably mitigates this problem and achieves competitive results compared to more complex unsupervised methods.
|
