Semantic-Aware Zero-Shot Multi-Label Classification for Remote Sensing Images
Journal of Intelligent Decision Making and Information Science · 03.10.2026 · 3 (16s) · 1499-1522
DOI: 10.17762/jidmis.v3.5066 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Intelligent Decision Making and Information Science, online erschienen am 03.10.2026. Verfasst von Vu Dinh Minh. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Remote sensing images often contain multiple land-cover and land-use concepts, making multi-label classification important for comprehensive scene understanding. However, conventional approaches typically require task-specific annotated data and retraining, which limits their flexibility when target labels or datasets change. This study proposes a lightweight semantic-aware zero-shot multi-label classification framework based on frozen vision-language models. CLIP and RemoteCLIP are used to generate image-label similarity scores, which are subsequently refined through a Semantic Label Affinity (SLA) mechanism derived from textual label representations. An Adaptive Thresholding (AT) strategy is then employed to determine image-specific multi-label decision boundaries without training an additional classifier. The framework requires neither backbone fine-tuning nor target-label co-occurrence statistics. Experiments on BigEarthNet-19 and MLRSNet demonstrate that the proposed components consistently improve zero-shot multi-label recognition. On BigEarthNet-19, the full CLIP-based configuration achieves 40.1% mAP and 58.3% Micro-F1, improving the corresponding zero-shot baseline by 9.9 and 14.4 percentage points, respectively. On MLRSNet, the RemoteCLIP-based configuration achieves 70.1% mAP and 77.1% Micro-F1, with gains of 7.6 and 9.5 percentage points. These results indicate that semantic label refinement and image-adaptive thresholding provide complementary benefits while preserving a training-free inference setting.
Zitieren
APA
Minh, V. D. (2026). Semantic-Aware Zero-Shot Multi-Label Classification for Remote Sensing Images. Journal of Intelligent Decision Making and Information Science, 3(16s), 1499-1522. https://doi.org/10.17762/jidmis.v3.5066
Vancouver
Minh VD. Semantic-Aware Zero-Shot Multi-Label Classification for Remote Sensing Images. Journal of Intelligent Decision Making and Information Science. 2026;3(16s):1499-1522. doi:10.17762/jidmis.v3.5066
BibTeX
@article{Minh2026,
author = {Minh, Vu Dinh},
title = {Semantic-Aware Zero-Shot Multi-Label Classification for Remote Sensing Images},
journal = {Journal of Intelligent Decision Making and Information Science},
year = {2026},
volume = {3},
number = {16s},
pages = {1499--1522},
doi = {10.17762/jidmis.v3.5066}
}
EUVIMED