EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

DINOv3-MSFNet: A Multi-Scale Feature Fusion Network for Land Cover Semantic Segmentation in Mountainous Forest Ecosystems

Xu Zhao, Lizhi Liu, Qiuliang Zhang

ISPRS International Journal of Geo-Information · 20.11.2026 · 15 (11) · 517

DOI: 10.3390/ijgi15110517 ↗

Kurz gefasst

Fachartikel in der Zeitschrift ISPRS International Journal of Geo-Information, online erschienen am 20.11.2026. Verfasst von Xu Zhao und 2 weiteren Autor:innen. Die Arbeit nennt 53 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Inner Mongolia Science and Technology Program, Inner Mongolia Natural Science Foundation for Young Scientists, Heilongjiang Post-doctoral Foundation.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Mountainous forest ecosystems are characterized by spatially interwoven land-cover types, substantial scale variation, interclass spectral similarity, and mixed boundaries, posing challenges to both global semantic modeling and local spatial-detail recovery. This study proposes DINOv3-MSFNet, a land-cover semantic segmentation network built upon a frozen vision foundation model. Hierarchical features are extracted from different Transformer depths of DINOv3-ViT-L/16. A Feature Enhancement Module (FEM) performs channel projection and spatial reorganization, a Cross-Scale Feature Fusion Module (CSFM) enables bidirectional full-scale interaction among hierarchical representations, and a Multi-Scale Fusion Decoder (MSFD) progressively reconstructs fine-grained land-cover structures. Experiments were conducted on two Sentinel-2 RGB datasets with Dynamic World pseudo-labels and the TALANDCOVER benchmark, using spatially independent splits and separate within-dataset training for TALANDCOVER. DINOv3-MSFNet achieved OA, mF1, and mIoU values of 92.34%, 82.59%, and 72.68%, respectively, on the Greater Khingan Mountains dataset; 92.34%, 86.13%, and 77.20% on the Lesser Khingan Mountains-related dataset; and 75.61%, 65.85%, and 51.54% on TALANDCOVER, yielding the best results among the evaluated methods. Four independent random runs further showed that the improvements in mF1 and mIoU were repeatable. Ablation experiments indicated that the MSFD provided the principal performance gain, while the FEM and CSFM played complementary roles through hierarchical feature adaptation and cross-scale interaction. Regional validation using manually interpreted reference samples showed that, after limited fine-tuning, the model effectively preserved the spatial continuity of extensive forest regions, although wetland and grassland remained the principal sources of uncertainty. These evaluations assess agreement with Dynamic World pseudo-labels, within-dataset TALANDCOVER performance, and regional performance after limited fine-tuning. They support robustness within datasets, but not direct cross-regional generalization. Overall, DINOv3-MSFNet effectively integrates global semantic modeling with local boundary refinement, offering a solution for fine-grained land-cover semantic segmentation in complex mountainous forest ecosystems.

Förderung

  • Inner Mongolia Science and Technology Program
  • Inner Mongolia Natural Science Foundation for Young Scientists
  • Heilongjiang Post-doctoral Foundation
  • Natural Science Foundation of Xinjiang Uygur Autonomous Region
  • Inner Mongolia Autonomous Region Science and Technology "Breakthrough" Project

Zitieren

APA

Zhao, X., Liu, L., & Zhang, Q. (2026). DINOv3-MSFNet: A Multi-Scale Feature Fusion Network for Land Cover Semantic Segmentation in Mountainous Forest Ecosystems. ISPRS International Journal of Geo-Information, 15(11), 517. https://doi.org/10.3390/ijgi15110517

Vancouver

Zhao X, Liu L, Zhang Q. DINOv3-MSFNet: A Multi-Scale Feature Fusion Network for Land Cover Semantic Segmentation in Mountainous Forest Ecosystems. ISPRS International Journal of Geo-Information. 2026;15(11):517. doi:10.3390/ijgi15110517

BibTeX
@article{Zhao2026,
  author = {Zhao, Xu and Liu, Lizhi and Zhang, Qiuliang},
  title = {DINOv3-MSFNet: A Multi-Scale Feature Fusion Network for Land Cover Semantic Segmentation in Mountainous Forest Ecosystems},
  journal = {ISPRS International Journal of Geo-Information},
  year = {2026},
  volume = {15},
  number = {11},
  pages = {517},
  doi = {10.3390/ijgi15110517}
}
Kostenlos starten →