GroundSet: A Cadastral-Grounded Dataset for Spatial Understanding with Vector Data

A new dataset called GroundSet aims to sharpen how artificial intelligence models interpret satellite and aerial imagery by grounding their understanding in official land-registry maps. The resource, detailed in a paper submitted in 2026, pairs 510,000 high-resolution images with 3.8 million annotated objects across 135 semantic categories [1][2]. The dataset is built on verifiable cadastral vector data, a type of official map record that precisely delineates property boundaries and land-use parcels [1][2]. In geographic information science, such features are formal representations of real-world phenomena, defined by their spatial location, form, and characteristics [3]. By tying image analysis directly to these authoritative records, the researchers sought to overcome a persistent weakness in current multimodal large language models, which often fail at fine-grained spatial reasoning when applied to remote sensing tasks [2]. Earth observation satellites, which have been in operation since the Soviet Union launched Sputnik 1 in 1957, now number nearly 13,000 in orbit and gather information for reconnaissance, mapping, and environmental monitoring [5]. The imagery they produce is vast, but translating it into actionable insights for urban planning or disaster management requires models that can accurately identify and locate specific features [2]. The authors of the GroundSet paper found that even specialized remote-sensing models and commercial systems, such as Gemini, perform poorly in zero-shot settings where they must interpret scenes without prior task-specific training [2]. To establish a performance baseline, the team conducted an instruction-tuning benchmark covering seven spatial reasoning tasks using a standard LLaVA architecture [2]. The results indicated that high-fidelity supervision from the cadastral-grounded dataset allowed the standard architecture to master fine-grained spatial grounding without complex modifications [2]. The paper, authored by Roger Ferrod, was submitted to the arXiv preprint server on March 15, 2026, and last revised on June 24, 2026 [1].