GeoMapCLIP, a Fine-Tuned GeoCLIP to Geolocate Satellite Imagery
August 20, 2025 11:00 am (Central Time)
Abstract
This talk presents our work on a geospatial vision system, GeoMapCLIP, for the I-GUIDE AI Spatial Challenge. The project focuses on developing AI models that can interpret geospatial map images and accurately extract coordinates, thereby enabling the automated understanding of map content, including scale, symbols, and location. Building on the CLIP and GeoCLIP frameworks, we adapted the model for satellite maps by training it with ArcGIS imagery. While GeoCLIP performs well on social media and popular location images, it falls short on map-specific tasks. Our fine-tuned GeoMapCLIP significantly improves performance in this domain, demonstrating better localization accuracy on unfamiliar satellite maps. In this seminar, we’ll share our model design process, dataset selection strategy, and fine-tuning methodology. We’ll also showcase comparative results that highlight the strengths of GeoMapCLIP in geospatial AI tasks. This work lays the foundation for integrating spatial reasoning into broader language-vision systems and supporting the use of legacy geospatial data in AI workflows.
Speakers
Jungha Woo
Rosen Center for Advanced Computing, Purdue University
Jungha Woo is a Software Engineer in the Research Computing at the Purdue University. His Ph.D. work included analyzing investors’ behavioral biases in the U.S. stock markets and implementing profitable strategies utilizing irrational behaviors. His experience and interests lie in the statistical analysis of scientific data, and software development. Jungha develops scientific software to help high-performance computational communities run models and predict execution time of jobs.
Elham Barezi
Rosen Center for Advanced Computing, Purdue University
Dr. Elham Barezi is a Lead AI research scientist at Purdue's Rosen Center for Advanced Computing. Prior, she worked as research staff for Michigan State University, and as a project manager at the Center for AI Research at the Hong Kong University of Science and Technology. She is experienced in the theory of machine learning, multimodal learning, and natural language processing. She has done projects in knowledge base analysis, visual question answering, multimodal therapy and personality recognition, multi-task and multi-label learning, and investigating compression and sampling methods for large-scale machine learning and deep neural networks.