This poster presents a framework for on-demand Earth observation data cubes that enables semantic, knowledge-based querying of multimodal EO data. Using STAC metadata, data are fetched on demand and organised into regular space–time cubes. A dedicated semantic query language allows domain knowledge and concepts such as “forest”, “disturbance”, or “clouds” to be encoded explicitly in reusable models. The framework is implemented as a standalone Python library (gsemantique) that can be deployed both locally and in the cloud. Chunking and parallelisation support efficient mesoscale analyses. An application example demonstrates the knowledge-based assessment of forest disturbances, combining multimodal and multitemporal EO data with expert knowledge in a transparent way. The poster was presented at the ESA Living Planet Symposium 2025 in Vienna, Austria