As climate change fuels increasingly severe weather events like floods, hurricanes, droughts, and wildfires, traditional disaster response methods are struggling to keep up. While advances in satellite technology, drones, and remote sensors allow for better monitoring, access to this vital data remains limited to a few organizations, leaving many researchers and innovators without the tools they need. The flood of geospatial data being generated daily has also become a challenge—overwhelming organizations and making it harder to extract meaningful insights. To address these issues, scalable, accessible, and intelligent tools are needed to turn vast datasets into actionable climate insights. This is where geospatial AI becomes vital—an emerging technology that has the potential to analyze large volumes of data, providing more accurate, proactive, and timely predictions. This article explores the groundbreaking collaboration between IBM and NASA to develop advanced, more accessible geospatial AI, empowering a wider audience with the tools necessary to drive innovative environmental and climate solutions.
Why IBM and NASA Are Pioneering Foundation Geospatial AI
Foundation models (FMs) represent a new frontier in AI, designed to learn from vast amounts of unlabeled data and apply their insights across multiple domains. This approach offers several key advantages. Unlike traditional AI models, FMs don’t rely on massive, painstakingly curated datasets. Instead, they can finetune on smaller data samples, saving both time and resources. This makes them a powerful tool for accelerating climate research, where gathering large datasets can be costly and time-consuming.
Moreover, FMs streamline the development of specialized applications, reducing redundant efforts. For example, once an FM is trained, it can be adapted to several downstream applications such as monitoring natural disasters or tracking land use without requiring extensive retraining. Though the initial training process can demand significant computational power, requiring tens of thousands of GPU hours. However, once they are trained, running them during inference takes mere minutes or even seconds.
Additionally, FMs could make advanced weather models accessible to a wider audience. Previously, only well-funded institutions with the resources to support complex infrastructure could run these models. However, with the rise of pre-trained FMs, climate modeling is now within reach for a broader group of researchers and innovators, opening up new avenues for faster discoveries and innovative environmental solutions.
The Genesis of Foundation Geospatial AI
The vast potential of FMs has led IBM and NASA to collaborate for building a comprehensive FM of the Earth’s environment. The key objective of this partnership is to empower researchers to extract insights from NASA’s extensive Earth datasets in a manner that is both effective and accessible.
In this pursuit, they achieve a significant breakthrough in August 2023 with the unveiling of a pioneering FM for geospatial data. This model was trained on NASA’s vast satellite dataset, comprising a 40-year archive of images from the Harmonized Landsat Sentinel-2 (HLS) program. It uses advanced AI techniques, including transformer architectures, to efficiently process substantial volumes of geospatial data. Developed using IBM’s Cloud Vela supercomputer and the watsonx FM stack, the HLS model can analyze data up to four times faster than traditional deep learning models while requiring significantly fewer labeled datasets for training.
The potential applications of this model are extensive, ranging from monitoring land use changes and natural disasters to predicting crop yields. Importantly, this powerful tool is freely available on Hugging Face, allowing researchers and innovators worldwide to utilize its capabilities and contribute to the advancement of climate and environmental science.
Advances in Foundation Geospatial AI
Building on this momentum, IBM and NASA have recently introduced another groundbreaking open-source model FM: Prithvi WxC. This model is designed to address both short-term weather challenges and long-term climate predictions. Pre-trained on 40 years of NASA’s Earth observation data from the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2), the FM offers significant advancements over traditional forecasting models.
The model is built using a vision transformer and a masked autoencoder, enabling it to encode spatial data over time. By incorporating a temporal attention mechanism, the FM can analyze MERRA-2 reanalysis data, which integrates various observational streams. The model can operate on both a spherical surface, like traditional climate models, and a flat, rectangular grid, allowing it to change between global and regional views without losing resolution.
This unique architecture enables the Prithvi to be fine-tuned across global, regional, and local scales, while running on a standard desktop computer in seconds. This FM model can be employed for a range of applications including forecasting local weather to predicting extreme weather events, enhancing the spatial resolution of global climate simulations, and refining the representation of physical processes in conventional models. Additionally, Prithvi comes with two fine-tuned versions designed for specific scientific and industrial uses, providing even greater precision for environmental analysis. The model is freely available on hugging face.
The Bottom Line
IBM and NASA’s partnership is redefining geospatial AI, making it easier for researchers and innovators to address pressing climate challenges. By developing foundation models that can effectively analyze large datasets, this collaboration enhances our ability to predict and manage severe weather events. More importantly, it opens the door for a wider audience to access these powerful tools, previously limited to well-resourced institutions. As these advanced AI models become accessible to more people, they pave the way for innovative solutions that can help us respond to climate change more effectively and responsibly.
