.jpg)
The interpretation of satellite images has advanced remarkably in recent years, but a complex challenge remains: identifying specific objects from natural language descriptions that include spatial, comparative, and ordinal relationships. For example, expressions like 'the largest ship in the port' or 'the second tennis court from the left' require a level of geometric and semantic reasoning that traditional segmentation models cannot accurately capture. Until now, training-free solutions relied on implicit activations or similarities between regions and text, offering very limited control over these dominant relationships in the aerial domain.
In this context, GeoSelect emerges, an innovative pipeline that rethinks referential segmentation as the execution of a typed spatial program. Instead of relying on neural networks trained with labeled data, GeoSelect uses a frozen language model (without fine-tuning) to synthesize the natural expression into a small domain-specific language. A well-formedness verifier accepts the program, and a deterministic executor carries it out. The key lies in a central abstraction: a scored candidate set type under which all operators are composed. Continuous geometric fields materialize position and proximity as dense pixel-level maps, while discrete set and order operators add extreme, ordinal, counted union, and relational constructs that fields alone cannot express. Since execution is explicit, every intermediate program, field, and ranking is inspectable, and a reliability ladder degrades any failed program to a special fields-only case, ensuring every expression returns a response.
The results are compelling: GeoSelect achieves 58.86 mIoU on RRSIS-D test and 55.27 mIoU on RISBench test, more than double the best previous training-free method on RRSIS-D, without any referential supervision and using a single GPU. A controlled analysis attributes the gain to explicit execution, not the backbone; an oracular decomposition locates the residual gap in detection recall on RRSIS-D and selection on RISBench, and an exposure audit confirms robustness against pretraining leakage. This demonstrates that a spatial program-based approach can far surpass purely connectionist approaches when it comes to following complex spatial instructions.
For companies working with geospatial data, this advancement opens up very interesting possibilities. Integrating GeoSelect into workflows for precision agriculture, logistics, defense, or urban planning allows querying satellite images with natural language questions and obtaining precise answers, without the need for costly labeled datasets or massive training infrastructure. At Q2BSTUDIO, we develop custom applications that incorporate cutting-edge artificial intelligence, and we know that solutions like GeoSelect fit perfectly into production environments where interpretability and efficiency are critical. Additionally, we combine these capabilities with AWS and Azure cloud services to scale the processing of large volumes of images, ensuring availability and performance.
AI for businesses is not limited to deep learning models; it also encompasses the creation of AI agents that autonomously execute spatial tasks, such as detecting changes in crops or monitoring infrastructure. Cybersecurity is equally relevant when handling sensitive location data; therefore, we offer cybersecurity and pentesting services to protect both systems and data. And once results are generated, Power BI and other business intelligence services allow for interactive visualization and analysis of geospatial information, facilitating strategic decision-making.
GeoSelect represents a paradigm shift in satellite referential segmentation, demonstrating that combining language models with programmatic execution can achieve superior control and precision without the need for additional training. At Q2BSTUDIO, we are prepared to help organizations adopt such technologies, integrating custom software, advanced artificial intelligence, and cloud solutions to turn satellite images into actionable knowledge. The invitation is to explore how these tools can transform your analytical and operational processes, always with a practical and secure approach.
