Recent advances in remote sensing and machine learning are enabling a new generation of geospatial understanding, driven by large-scale, general-purpose foundation models. This session highlights progress beyond task-specific models, focusing on design, pretraining, and fine-tuning of geospatial foundation models leveraging diverse data sources such as optical, SAR, and hyperspectral imagery. We welcome contributions that explore spatiotemporal representation learning, domain adaptation, and downstream applications in areas like land cover classification, disaster monitoring, and environmental change detection. Submissions on knowledge distillation and model compression are also encouraged—particularly efforts that adapt open-source foundation models into lightweight, localized systems for resource-constrained or regional use cases. We are especially interested in work advancing geospatial reasoning and multimodal understanding within vision-language frameworks, toward models that can both perceive and interpret the complexities of Earth systems. By bridging scalable AI and Earth observation, this session supports the development of accessible, intelligent, and transparent tools for geoscientific research.