Domain-Adapted Remote Sensing for Urban Change Detection Using Weak Supervision from Maps

Abstract

The proliferation of high-resolution Earth observation data has created unprecedented opportunities for monitoring urban dynamics. However, the efficacy of supervised deep learning models for change detection is frequently curtailed by the scarcity of pixel-level ground truth annotations and the statistical heterogeneity inherent in multi-temporal or cross-sensor imagery. This paper presents a novel framework for Domain-Adapted Remote Sensing for Urban Change Detection (DARS-UCD), which leverages weak supervision from openly available cartographic data, specifically OpenStreetMap (OSM). We propose a dual-stream architecture that harmonizes feature representations between a labeled source domain and an unlabeled target domain, where the target domain supervision is derived solely from noisy, outdated, or incomplete map rasters. To mitigate the domain shift and label noise, we introduce a Map-Guided Uncertainty Weighting (MGUW) mechanism coupled with an adversarial domain adaptation module. Extensive experiments demonstrate that our approach significantly outperforms standard unsupervised methods and achieves competitive performance relative to fully supervised baselines. The results validate the utility of integrating semantic map priors into remote sensing pipelines, offering a scalable solution for global-scale urban monitoring.