GeoVLM: Geography-Aware Multimodal Visual Reasoning for Earth Science Applications
Year: 2026
Type: preprint
Abstract: The integration of large vision-language models with earth science data presents unique challenges including the scarcity of labeled geoscientific data, the need for effective long-context reasoning over temporal satellite imagery, and the heterogeneity of multimodal earth observation sources. In this paper, we propose GeoVLM, a unified multimodal visual reasoning framework specifically designed for earth science applications. GeoVLM introduces three key innovations: a Geography-Aware Visual In-Context Learning module that selects demonstration examples based on spatial, spectral, and temporal similarity for few-shot geospatial interpretation; a Temporal Redundancy-Aware Attention mechanism that adaptively suppresses visual redundancy in long satellite image sequences while preserving critical change information; and a Modular Multi-Agent Cross-Modal Fusion architecture that enables collaborative reasoning across heterogeneous data sources including optical imagery, synthetic aperture radar, meteorological measurements, and geological surveys. We evaluate GeoVLM on six benchmark datasets spanning three core tasks: few-shot geospatial interpretation on DFC15 and Potsdam, temporal change detection on LEVIR-CD and WHU-CD, and extreme event reasoning on ClimateBench and MultiEarth. Experimental results demonstrate that GeoVLM achieves state-of-the-art performance across all tasks, with improvements of 3.5 percent mIoU on DFC15, 2.7 percent F1 on LEVIR-CD, and 0.47 RMSE reduction on ClimateBench compared to the best existing methods. Ablation studies confirm the complementary contributions of each module, and human evaluation by remote sensing experts further validates the quality of our predictions.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7163692011