How IBM and NASA Are Redefining Geospatial AI to Tackle Climate Challenges

Like climate change fuels With increasingly severe weather events such as floods, hurricanes, droughts and wildfires, traditional disaster response methods are struggling to keep pace. While advances in satellite technology, drones and remote sensors enable 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 generated every day has also become a challenge, overwhelming organizations and making it more difficult to gain meaningful insights. Addressing these issues requires scalable, accessible, and intelligent tools to transform massive data sets into actionable climate insights. This is true geospatial AI is becoming critical: an emerging technology that has the potential to analyze large amounts of data, allowing for more accurate, proactive, and timely predictions. This article explores the groundbreaking collaboration between IBM and NASA to develop advanced, more accessible geospatial AI, giving a broader audience the tools needed to drive innovative environmental and climate solutions.

Why IBM and NASA are pioneering Geospatial AI

Foundation models (FMs) represent a new frontier in AI, designed to learn from massive amounts of unlabeled data and apply their insights across multiple domains. This approach offers several important advantages. Unlike traditional AI models, FMs do not rely on huge, carefully curated data sets. Instead, they can refine smaller data samples, saving both time and resources. This makes them a powerful tool for accelerating climate research, where collecting large data sets can be expensive and time-consuming.

Additionally, FMs streamline the development of specialized applications, reducing redundant efforts. For example, once an FM is trained, it can be adapted to different downstream applications, such as monitoring natural disasters or tracking land use, without the need for extensive retraining. Although the initial training process can require significant computing power, requiring tens of thousands of GPU hours. However, once trained, performing them during inference takes only a few minutes or even seconds.

In addition, FMs can make advanced weather models accessible to a wider audience. Previously, only well-funded institutions with the resources to support complex infrastructure could use 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 new avenues for faster discoveries and innovative environmental solutions.

The Genesis of Foundation Geospatial AI

The enormous potential of FMs has led IBM and NASA to work together to build a comprehensive FM of the Earth’s environment. The main goal of this partnership is to enable researchers to gain insights from NASA’s extensive Earth data sets in a way that is both effective and accessible.

In this endeavor, they will achieve a major breakthrough in August 2023 with the unveiling of a groundbreaking product FM for geospatial data. This model is trained on NASA’s massive satellite dataset, which includes a 40-year archive of NASA images Harmonized Landsat Sentinel-2 (HLS) program. It uses advanced AI techniques, including transformer architectures, to efficiently process significant amounts of geospatial data. Developed with the help of 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 available for free available on Hugging Face, allowing researchers and innovators around the world to harness its potential and contribute to the advancement of climate and environmental sciences.

Advances in Foundation Geospatial AI

Building on this momentum, IBM and NASA 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 forecasts. 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 improvements over traditional forecasting models.

The model was built using a vision transformer and a masked autoencoderallowing it to encode spatial data over time. By recording a temporal attention mechanismcan analyze the FM MERRA-2 reanalysis data, integrating different observation streams. The model can operate on either a spherical surface, like traditional climate models, or on a flat, rectangular grid, allowing it to switch between global and regional views without loss of resolution.

This unique architecture allows Prithvi to be refined on a global, regional and local scale, while running in seconds on a standard desktop computer. This FM model can be used for a range of applications, including predicting local weather to predicting extreme weather events, improving the spatial resolution of global climate simulations, and refining the representation of physical processes in conventional models. Moreover, Prithvi comes with two refined versions designed for specific scientific and industrial use, offering even greater accuracy for environmental analysis. The model is free available on hugging face.

The bottom line

The collaboration between IBM and NASA is redefining geospatial AI, making it easier for researchers and innovators to tackle pressing climate challenges. By developing baseline models that can effectively analyze large data sets, this collaboration increases our ability to predict and manage severe weather events. More importantly, it opens the door for a wider audience to access these powerful tools, which were 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.