Skip to content
Clark University
Center for
Geospatial Analytics

Prithvi Foundation Model (Collaboration with NASA and IBM)

Background

Foundation Models (FMs) are revolutionizing machine learning (ML) by harnessing self-supervised techniques on vast unlabeled datasets, such as global satellite imagery. This innovative approach involves training FMs on extensive datasets and refining them for specific applications with minimal labeled data, proving highly effective in terms of accuracy and training cost. With satellite missions expected to generate 250,000 terabytes of data by 2024, the need for efficient data analysis becomes paramount. Dr. Hamed Alemohammad, Director of Clark Center for Geospatial Analytics, in collaboration with NASA and IBM, is pioneering the world’s first geospatial AI FM.

The development and advancement of FMs have initiated a paradigm shift in machine learning (ML) model construction. These models, trained on large-scale input data in a self-supervised framework, offer adaptability for downstream tasks using a limited sample of training data. To evaluate the value of FMs in Earth Observation (EO) applications, a collaborative team from NASA and IBM has developed Prithvi FM model utilizing Harmonized Landsat Sentinel (HLS) multi-spectral data. This project aims to assess the efficacy and potential of FMs in advancing the field of geospatial analytics.

foundation model graphic

Impact and Innovation

The geospatial AI FM, a breakthrough in generative AI for geospatial analytics, offers a unique approach to derive insights from large geospatial datasets. Similar to constructing a building foundation before customization, this FM expedites the development of generative AI models for various applications, saving time and resources. The project has demonstrated significant improvement in mapping floods and burn scars, showcasing its potential impact.

Clark CGA team has implemented two downstream applications for cloud gap filling and crop and land cover classification using multi-temporal data. The cloud gap filling application shows that Prithvi can predict multi-spectral cloudy pixels across complex geographies with high accuracy. Fine-tuning Prithvi with varying sample sizes shows that it consistently performs better compared to a baseline Conditional GAN model, and it better preserves the relationship between multispectral bands.

By open-sourcing the FM and the codebase, the team aligns with NASA’s Open-Source Science Initiative, aiming to transform the analysis of observational data and enhance our understanding of the planet. This collaboration between Clark University, NASA, and IBM has the potential to revolutionize how geospatial analytics is approached globally.

Resources