The rapid growth of Earth Observation data volumes has surpassed the capacity of conventional supervised learning pipelines, prompting a shift toward foundation models that rely on self-supervised pre-training to support a wide range of downstream applications. This paper provides an experimental review of state-of the-art GeoAI foundation models, examining their performance in zero-shot scenarios for semantic segmentation, scene classification, and object detection. A multi-source dataset incorporating Sentinel-2, PlanetScope, and very high resolution aerial imagery over Galway city, Ireland is employed to evaluate models such as Prithvi, RemoteCLIP, SAMGeo based architectures, and lightweight vision language models. Particular emphasis is placed on the transformation of raw imagery into analysis ready vector products within a reproducible, desktop scale workflow. The experimental analysis shows that foundation models consistently achieve strong performance in detecting broad land cover classes, including vegetation and water bodies, across varying spatial resolutions. In zero-shot semantic segmentation tasks, Segment Anything Model (SAM) based approaches exhibited high geometric fidelity when applied to very high resolution aerial data; however, their performance degraded noticeably on medium resolution satellite imagery, where fine scale features such as vehicles and small buildings could not be reliably resolved. Scene classification was effective in differentiating urban and natural environments, although its robustness depended heavily on prompt design and sensor specific spectral properties. Overall, the results highlight that although foundation models substantially reduce dependence on labeled training data, they remain constrained by insufficient topological coherence and limited exploitation of spectral information for fully automated, high accuracy cartographic production. The current experiments does not involve any model fine-tuning; all experiments are conducted using zero-shot inference with pre-trained models. Future work should prioritize the development of multimodal architectures that explicitly incorporate multispectral and temporal information beyond RGB focused representations.

抄録全体を表示