Self-supervised Change Detection via Cooperative Learning: A Two-Player Model

Change detection (CD) is a critical task in remote sensing, especially for monitoring dynamic urban environments. In this work, we present a novel, self-supervised CD approach that leverages a two-player architecture to improve the reliability and accuracy of change predictions. The 2Player framework consists of two cooperating models: Player 1, a change detection model, and Player 2, a supporting autoencoder trained for reconstruction. The two models work in tandem: after reconstructing the second image from the first one, the second player highlights areas of potential change where reconstruction is difficult. On the other hand, the first player focuses on predicting changes based on these indications and shares its change maps with the autoencoder, enabling it to disregard these areas during reconstruction. This mutual guidance creates a robust self-supervised mechanism that we test with a FC-Siam-Diff model for the first player and a standard UNet for the second. We evaluate the 2Player framework on a refined version of the HRSCD dataset, where labels have been enhanced using IGN’s BD TOPO® model, resulting in a cleaner dataset of 10,000 images. Experimental results show that our approach achieves improved performances in detecting changes, outperforming traditional unsupervised baselines.