Anaximander: Interactive Orchestration and Evaluation of Geospatial Foundation Models

Geospatial Foundation Models – deep learning architectures trained on large-scale Earth observation and remote sensing data – have proved useful for object detection, semantic segmentation, and other prediction tasks on geospatial and geographic data. However, experts in domains most likely to benefit from applying these models are accustomed to traditional geographic information system (GIS) workflows, largely driven by graphical user interfaces, with limited code-based workflows typical of those required to deploy deep learning models. To bridge this gap, we have developed Anaximander: a system for integrating state-of-the-art foundation models into traditional GIS and cartographic workflows. This project, consisting of a model orchestration server and a plugin for QGIS (a commonly-used, open-source GIS tool), allows analysts to load models from a variety of sources, orchestrate embedding generation on either a local or remote GPU instance, and refine the model outputs for downstream, problem-specific tasks. We will demonstrate the system capabilities in the context of an active project involving croptype mapping in Kenya. Learn more about Microsoft Research Lab – Africa, Nairobi: https://www.microsoft.com/en-us/research/lab/microsoft-research-lab-africa-nairobi/seminars/