Scholarlink

<p>Geospatial foundation models (GFMs) are fundamentally restructuring land-use and land-cover (LULC) classification, yet no review has empirically traced this transformation over time. This study employs a yearly topic network analysis on a near-exhaustive corpus of 30 Web of Science publications (2022–2025) at the GFM–LULC intersection, constructing keyword co-occurrence networks from 156 author keywords mapped to eight thematic categories and calculating composite hub-node centrality scores for each annual window. The hub-node trajectory migrated from Deep Learning &amp; AI (2022, H = 1.000) through Land Use / Land Cover (2023, H = 0.929) to Environmental Applications (2024–2025, H = 0.786–0.929), providing empirical evidence that GFMs have redirected the field’s structural center from methodology toward applied domains. This paradigmatic transition proceeds through four phases: task-specific consolidation, infrastructural readiness, disruption of the vision foundation model, and application-centric integration. Five dynamics underpin this restructuring, spanning the displacement of task-specific learning by representation-centric approaches, reduced dependence on labeled data, multimodal sensor convergence, predictive spatio-temporal modeling, and application-domain diversification. Sensitivity analysis across nine weighting configurations and permutation testing (<i>p</i> &lt; 0.001) confirms the trajectory’s robustness. Emergent frontiers include domain-native GFMs, physics-informed hybridization, GFM–LLM convergence, real-time monitoring, and equity governance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Transformative dynamics of geospatial foundation models in land use and land cover classification

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Young-Seok Hwang,
  • Minjae Kim,
  • JuSeok Lee,
  • Shin-Ae Park,
  • Aetti Kang

摘要

Geospatial foundation models (GFMs) are fundamentally restructuring land-use and land-cover (LULC) classification, yet no review has empirically traced this transformation over time. This study employs a yearly topic network analysis on a near-exhaustive corpus of 30 Web of Science publications (2022–2025) at the GFM–LULC intersection, constructing keyword co-occurrence networks from 156 author keywords mapped to eight thematic categories and calculating composite hub-node centrality scores for each annual window. The hub-node trajectory migrated from Deep Learning & AI (2022, H = 1.000) through Land Use / Land Cover (2023, H = 0.929) to Environmental Applications (2024–2025, H = 0.786–0.929), providing empirical evidence that GFMs have redirected the field’s structural center from methodology toward applied domains. This paradigmatic transition proceeds through four phases: task-specific consolidation, infrastructural readiness, disruption of the vision foundation model, and application-centric integration. Five dynamics underpin this restructuring, spanning the displacement of task-specific learning by representation-centric approaches, reduced dependence on labeled data, multimodal sensor convergence, predictive spatio-temporal modeling, and application-domain diversification. Sensitivity analysis across nine weighting configurations and permutation testing (p < 0.001) confirms the trajectory’s robustness. Emergent frontiers include domain-native GFMs, physics-informed hybridization, GFM–LLM convergence, real-time monitoring, and equity governance.