EUVIMEDEUVIMED entdecken
Crossref

Spatial Location‐Constrained Fine‐Tuning of Vision‐Language Models for Urban Street‐Scene Feature Recognition

Sheng Hu, Wenchen Li, Yifan Zhang, Yongyang Xu, Run Wang, Liang Wu, Zhong Xie

Transactions in GIS · 2026-11 · 30 (7)

DOI: 10.1111/tgis.70464 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Transactions in GIS, online erschienen am 16.11.2026. Verfasst von Sheng Hu und 6 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Basic and Applied Basic Research Foundation of Guangdong Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Vision‐language models (VLMs) have shown remarkable capabilities in general image understanding, yet their direct application to urban street‐scene feature recognition faces two key challenges. First, these models are generally trained solely on image‐text data, which inherently lacks geographic prior knowledge, such as spatial proximity and regional heterogeneity, leading to spatially inconsistent predictions. Second, full‐parameter fine‐tuning demands substantial computational resources, placing it beyond the reach of most GIScience researchers. To address these issues, we propose Geo‐VLFT, a spatial location‐constrained method for fine‐tuning VLMs. Building on the Low‐Rank Adaptation (LoRA) framework, Geo‐VLFT introduces a spatial proximity consistency regularization term that encourages the model to produce similar latent representations for spatially adjacent street‐view images, guided by the First Law of Geography that nearby locations tend to share similar characteristics. This regularization steers the model's optimization to incorporate spatial proximity priors into the fine‐tuning process. Experiments on a multi‐city street‐view dataset spanning five functional categories (commercial, residential, industrial, public service, and urban village) demonstrate that Geo‐VLFT substantially outperforms standard LoRA fine‐tuning in both classification accuracy and spatial consistency. Ablation studies further confirm the geographic rationale behind key design choices, including Gaussian distance weighting and an attraction‐only (no repulsion) constraint. The method trains approximately 8.4 million trainable parameters (about 0.12% of the 7.0 billion base‐model parameters) and can run efficiently on a single consumer‐grade GPU, substantially lowering the barrier to deploying large models for geographic applications. Geo‐VLFT offers a reproducible, lightweight paradigm for embedding spatial constraints into VLMs fine‐tuning, with potential relevance to applications such as urban health assessment and urban spatial structure identification.

Förderung

  • National Natural Science Foundation of China
  • Basic and Applied Basic Research Foundation of Guangdong Province

Zitieren

APA

Hu, S., Li, W., Zhang, Y., Xu, Y., Wang, R., Wu, L., & Xie, Z. (2026). Spatial Location‐Constrained Fine‐Tuning of Vision‐Language Models for Urban Street‐Scene Feature Recognition. Transactions in GIS, 30(7). https://doi.org/10.1111/tgis.70464

Vancouver

Hu S, Li W, Zhang Y, Xu Y, Wang R, Wu L, et al. Spatial Location‐Constrained Fine‐Tuning of Vision‐Language Models for Urban Street‐Scene Feature Recognition. Transactions in GIS. 2026;30(7). doi:10.1111/tgis.70464

BibTeX
@article{Hu2026,
  author = {Hu, Sheng and Li, Wenchen and Zhang, Yifan and Xu, Yongyang and Wang, Run and Wu, Liang and Xie, Zhong},
  title = {Spatial Location‐Constrained Fine‐Tuning of Vision‐Language Models for Urban Street‐Scene Feature Recognition},
  journal = {Transactions in GIS},
  year = {2026},
  volume = {30},
  number = {7},
  doi = {10.1111/tgis.70464}
}
Kostenlos starten →