Nenadzorovana detekcija sprememb v satelitskih posnetkih z generiranjem sintetičnih primerov

Smole, Enej (author), Čehovin Zajc, Luka (mentor), Rolih, Blaž (comentor)

Detekcija sprememb v daljinsko zaznanih satelitskih posnetkih navadno zahteva velike količine ročno označenih parov slik, kar je drago in časovno potratno. Diplomska naloga za osnovo uporabi ogrodje BTC s kodirnikom DINOv3 in model nauči izključno na sintetičnih spremembah, pri čemer podrobneje razišče DRAEM-u podoben pristop z organsko obliko maske in mehkim mešanjem vsebine. Rezultati na treh podatkovnih zbirkah satelitskih posnetkov pokažejo, da preprostejše metode vodijo v učenje po bližnjici, medtem ko DRAEM-u podoben pristop ta problem občutno omili in doseže konkurenčne rezultate v primerjavi z bolj kompleksnimi nenadzorovanimi metodami.

• Language: Slovenian
• Keywords: nenadzorovano učenje, detekcija sprememb, DINOv3, sintetične spremembe, DRAEM, daljinsko zaznavanje
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• PID: 20.500.12556/RUL-187771
• Publication date in RUL: 14.09.2026

Secondary language

• Language: English
• Title: Unsupervised change detection in satellite images using synthetic example generation

Abstract

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.

• Keywords: unsupervised learning, change detection, DINOv3, synthetic changes, DRAEM, remote sensing