Environmental analysis made easier with IBM’s Geospatial Studio

The success of foundation models for language and code has paved the way for flexible, generalizable AI for just about any domain. This includes Earth’s climate, where advanced AI vision transformers can now handle the intricacies of geospatial data from sources like Earth-observing satellites and long-term climate simulations.

IBM Research has been collaborating with NASA to develop a family of geospatial foundation models aimed at analyzing Earth’s changing climate and the downstream effects. The first model in our Prithvi series has been downloaded more than 9,000 times since its open-source release last August, and has been used in summer schools and college programs around the world.

The newest member of the Prithvi family is a foundation model for weather and climate that IBM Research developed in collaboration with NASA and Oak Ridge National Laboratory. The transformer-based model has been pre-trained on NASA's MERRA-2 re-analysis dataset to learn atmospheric physics while dealing with missing information, and features a hierarchical attention mechanism to capture both short and long-range spatial dependencies.

It’s also the first grid-free foundation model for weather and climate, which means no assumptions have been made about the underlying data structure, allowing the model to scale from global to local areas without losing resolution.

The model’s flexible architecture makes it well-suited for a variety of applications. One that IBM Research is exploring is customized energy forecasts for wind farms. Tokens representing each wind turbine location are inserted into the model and integrated with global weather pattern knowledge the model learned during pre-training. If successful, the application could improve how renewable energy loads are assimilated into the grid.

The power of automation and visualization

This ability to peer into the future could reshape how industries manage energy loads, supply chain logistics, and natural disasters — among many other tasks. But one major barrier to customizing geospatial foundation models for different downstream tasks is fine-tuning.

Fine-tuning is how you customize a pre-trained foundation model to create practical applications for solving real-world problems. The Prithvi models, for example, can be fine-tuned to address a wide range of climate applications globally, using different datasets at different timescales.

Fine-tuning has traditionally required at least some expertise in curating data and training AI models. It can involve adapting the model’s architecture, cleaning and harmonizing data from different modalities, and incorporating data into the model without causing it to forget what it previously learned. In the case of a geospatial foundation model, fine-tuning may require integrating satellite images with historical weather observations, street-maps, and climate projections.

To make this process easier, IBM Research has developed the Geospatial Studio, a cloud-native toolkit that guides users through data curation, fine-tuning and model deployment, and visualization of results. The Studio largely automates data preparation for inference and fine-tuning.

It’s built on TerraTorch, an open-source framework library developed by IBM Research that lets you play with geospatial foundation models at varying levels of abstraction. TerraTorch configurations are provided for common fine-tuning tasks such as image segmentation and pixel-wise regression along with a selection of task-specific decoder heads.

The Studio can track and record each step in the training process, allowing you to monitor and measure the model’s accuracy and track each dataset that was used. The automated submission, scheduling, and tracking of fine-tuning tasks are integrated with tools for optimizing the model’s hyperparameters to find the best performing model configuration.

The graphical UI lets you visualize the model’s results, which is important for validating the model’s predictive skill and putting its results in context by overlaying additional information like asset locations or administrative boundaries.

IBM Research developed the Geospatial Studio to democratize access to geospatial foundation models and broaden their potential impact. IBM and NASA’s open-source models can be found in the Studio, along with other models developed by the research community. IBM Research recently used the Studio to tune its Granite geospatial models for forecasting changes in biomass (granite-geospatial-biomass) and wind strength (granite-geospatial-windforecasting).