Dual-domain representation alignment for unsupervised height estimation from cross-resolution remote sensing images

With the rapid advancement of remote sensing technology, height estimation from single-view remote sensing images has garnered significant attention due to its extensive applications in urban planning, environmental monitoring, and disaster management. Existing unsupervised domain adaptation methods for single-view remote sensing images often neglect the discrepancy in spatial resolution between the source and target domains. This oversight restricts their ability to generalize from low ground sample distance (GSD) to fine GSD images effectively. In this work, we explore a cross-resolution case encountered in many real-world applications to investigate the task of height estimation under unsupervised domain adaptation. This new context presents two formidable challenges: (1) How to capture resolution-invariant representations for better unsupervised domain adaptation; (2) How to maintain geometric integrity and spatial layout across domains. We propose a cross-resolution unsupervised height estimation framework with Dual-Domain Representation Alignment (DDRA) to address both challenges. The framework includes a dual-domain distillation strategy that enables the model to extract complementary knowledge from both fine GSD and coarse GSD domains without direct resolution matching, learning robust cross-domain representations through data augmentation and Transformer-based layer distillation. Additionally, a dual-stream alignment strategy is introduced for the unlabeled target inputs, involving contrastive learning to enhance structural awareness and a distillation objective to guide height estimation, thereby preserving geometric integrity and spatial layout across domains. We demonstrate that DDRA achieves state-of-the-art results on three challenging benchmarks: GAMUS, OsiDataset, and Enschede, while being effective for inputs of varying spatial resolution in both the source and target domains.