Abstract
<jats:p>Earth-science scene understanding increasingly relies on large, multi-modal Earth-observation (EO) archives that pair optical, multispectral, and SAR imagery with rich textual metadata. However, applying vision-language models (VLMs) to EO data is hindered by three challenges: retrieval pipelines typically ignore the textual query, pretraining is memory-and compute-bound at scale, and real-world inputs are often degraded by low-light conditions, sensor noise, or manipulation. We propose GeoFusion-VLM, an efficient vision-language model that addresses these issues through four tightly coupled components: (i) textual-prompt-aware retrieval, which conditions the visual encoder on the textual query to fetch the most relevant EO patch before reasoning; (ii) a synchronized cross-modal fusion block that aligns asynchronous optical and auxiliary streams via pace-aligned cross-attention; (iii) a memory-efficient large-batch optimizer that combines confidence-guided adaptive updates with a maximum-normalized element-wise ratio rule, cutting peak GPU memory from 41.6 GB to 24.8 GB; and (iv) a quality-aware preprocessing front-end that restores low-light and corrupted imagery prior to encoding. On GeoContext-100k, EO-Retrieval, and Surrogate-Phys, GeoFusion-VLM attains 83.5% scene-classification accuracy and 69.7% retrieval R@1, outperforming strong CLIP and LVLM baselines while remaining robust under textual-query noise, unseen sensors, and low-light degradation.</jats:p>