Blaž Rolih

Multispectral Earth observation models can now learn more effectively from optical vision foundation models using a dual-teacher distillation approach, boosting performance by up to 3.64% on key tasks.

By synthesizing changes directly in the latent space, MaSoN achieves state-of-the-art unsupervised change detection in remote sensing, generalizing to complex and rare scenarios where existing methods fall short.