Building a Scalable Platform for Geospatial AI Model Inference and Fine-Tuning

With the recent popularity of foundation models and their application in geospatial and climate domains, geospatial AI and machine learning workflows are becoming critical for research applications. We have observed scientists in domains such as biodiversity, land use, geophysical processes and natural disaster monitoring share bespoke notebooks and scripts to run their experiments. A need arises for automated, scalable, reproducible and sharable tools for operationalizing these experiments. We built a cloud-native geospatial platform that leverages foundation models pre-trained on climate and remote sensing data with two main components, supporting model fine-tuning and inference. The cloud-native architecture, built on Red Hat Openshift, ensures reproducibility, scalability, and consistency across fine-tuning and inference workflows. The goal of the platform is both to support non-expert users in harnessing the potential of geospatial foundation models, as well as to accelerate the work of experts. The fine-tuning service allows users to onboard fine-tuning datasets and create new application-specific models, including support for non-experts through templates and automation. Fine-tuning supports a range of model backbones including Terramind, Prithvi EO and other open-source models, through integration with TerraTorch, and facilitates hyperparameter optimization through TerraTorch Iterate. The inference service enables scalable inferencing against trained models (either pre-loaded, shared or user-tuned), intelligently parallelising requests across spatial and temporal domains, to ensure efficient use of available compute resources. To ease access to data for driving the pipelines, the platform includes integration with a geospatial AI data toolkit for querying (from a range of external data sources) and preparing data. The studio is accessed through a graphical web interface or through a set of RESTful APIs. A Python SDK provide powerful, programmatic access to the studio functionality and QGIS plugin enables traditional GIS users a familiar interface. Example applications based on the Prithvi and Terramind models include flood and wildfire burn scar mapping, land use classification, above ground biomass estimation and marine monitoring.