A Joint Spectral–Spatial Adapter for Domain Generalization in High-Resolution Remote Sensing Image Segmentation
IEEE Access · 2026 · 14 · 108435-108449
DOI: 10.1109/access.2026.3703898 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Yihui Liu und 4 weiteren Autor:innen. Die Arbeit nennt 79 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Shandong Provincial Natural Science Foundation.
Abstract (Originalsprache)
Remote sensing semantic segmentation is a core task in Earth observation and plays an important role in urban planning and land-use classification. The emergence of vision foundation model (VFM) has significantly boosted the segmentation performance. However, in real-world deployment, a pre-trained VFM must handle remote sensing images from other domains that it has never encountered. In this work, we advance domain generalization in remote sensing image segmentation by concentrating on the central challenge, namely, large cross-domain appearance shifts. A novel joint spectral-spatial adapter (SSA) is proposed to implement parameter-efficient fine-tuning on a frozen VFM, enabling better generalization capability. Technically, the proposed adapter consists of three components. Layer-wise spectral factorization combines a Haar-wavelet pyramid with spectrum-aware splitting to stabilize the content and structure-preserving details under cross-domain variations. Spectral-spatial mixture-of-expert modeling then refines these features through both spectrum and spatial updates, each of which is empowered by an expert model. Before decoding the segmentation map, energy-aware dynamic routing adaptively fuses expert outputs to suppress domain-specific styles without discarding boundaries and fine structures. On the ISPRS Vaihingen $\rightarrow $ Potsdam and Potsdam $\rightarrow $ Vaihingen settings, the proposed method achieves 62.6% and 71.0% mIoU, respectively, improving over the strongest comparable DINOv2-based baseline by 1.5 and 3.8 percentage points. On WHU Aerial $\rightarrow $ Satellite II and Satellite II $\rightarrow $ Aerial, it further improves F1/IoU to 46.9%/43.2% and 74.1%/72.3%, respectively.
Förderung
- Shandong Provincial Natural Science Foundation
Zitieren
APA
Liu, Y., Fu, Y., Li, C., Zhang, Z., & Cao, J. (2026). A Joint Spectral–Spatial Adapter for Domain Generalization in High-Resolution Remote Sensing Image Segmentation. IEEE Access, 14, 108435-108449. https://doi.org/10.1109/access.2026.3703898
Vancouver
Liu Y, Fu Y, Li C, Zhang Z, Cao J. A Joint Spectral–Spatial Adapter for Domain Generalization in High-Resolution Remote Sensing Image Segmentation. IEEE Access. 2026;14:108435-108449. doi:10.1109/access.2026.3703898
BibTeX
@article{Liu2026,
author = {Liu, Yihui and Fu, Yuning and Li, Chuanyao and Zhang, Zuoxuan and Cao, Jianfei},
title = {A Joint Spectral–Spatial Adapter for Domain Generalization in High-Resolution Remote Sensing Image Segmentation},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {108435--108449},
doi = {10.1109/access.2026.3703898}
}
EUVIMED