Introduction
Galileo is an open and highly multimodal source foundation model developed to process, analyze and understand various flows of earth observation data (EO) – including optical, radar, elevation, climate and auxiliaries – on a large scale. Galileo is developed with the support of researchers from McGill University, NASA Harvest AI2, Carleton University, the University of British Columbia, the Vector Institute and Arizona State University. Galileo aims to provide a unified general solution for critical applications such as agricultural land cartography, response to disasters and environmental surveillance.
Unlike previous remote sensing models limited to a single type of data or a scale, Galileo flexibly merges detection methods and is designed to recognize phenomena ranging from tiny objects (such as fishing boats, measuring only 1 to 2 pixels) with large characteristics changing slowly like glaciers.


Key characteristics and architecture
Multimodal transformer design
Galileo is based on a vision transformer architecture (VI), meticulously adapted to the process:
- Multispectral optical imaging (for example, Sentinel-2)
- Synthetic opening radar (SAR) (for example, Sentinel-1)
- Elevation and slope data (for example, NASA SRTM)
- Meteorological data / climate (for example, ERA5 precipitation and temperature)
- Terrestrial covers, population, night light and more
Flexible manipulation of entries:
Galileo's tokénisation pipeline divides remote sensing in spaces, schedules and groups of logical channels. This allows the model to process images, chronological series and static tabular data in a single architecture configuration.
Learning local and global functionality unified
A basic innovation is the Galileo self-supervised pre-training algorithm, which combines:
- Global losses: Encourage abstraction on large spatial or temporal contexts – identify “large” or slow change characteristics (glaciers, loss of forest).
- Local losses: Improve the sensitivity to tiny – crucial details to detect small rapidly evolving objects (boats, debris).
The local and global objectives differ in:
- Prediction depth: Global tasks target deep latent representations; Local tasks use shallow and linearly projected features.
- Masking strategies: Global tasks use structured and correlated space masks (forcing predictions over large intervals); Local tasks use random unstructured masks.
This double objective pre-training improves the representation of characteristics with several scales, which makes Galileo generalizable through tasks and robust even with limited labels.
Data set and pre-training strategy
To ensure semantic and geographic diversity, the Galileo pre-training data set covers the entire globe, sampled via a clustering approach to maximize the variety of land coverage and geographic spread. The data set includes more than 127,000 spatio-temporal aligned samples, each comprising four categories and nine types of remote sensing data.
Pre-training occurs for 500 eras on large calculation resources. Key aspects:
- Lot size: Effective size of the 512 lot.
- Data increases: Sizes of rotation, rotation and variable patch.
- Optimization: Adamw with the programmed learning rate and the decrease in weight sweeps.


Reference results
Higher generalization
Galileo is compared 11 diversified data sets and 15 downstream tasksCovering the classification of chronological image and pixel series, as well as segmentation. More specifically, he dominates on public data sets such as Eurosat, Bidearthnet, SO2SAT, Mados (Marine Debris), SEN1FLods11 (Cartography of Sar), Cropharvest (Multimodal Cultures Classification) and many others.
Prime performance of Galileo-Base (vital base):
- Classification (finetune):
- Eurosat: 97.7% (TOP-1 precision, 100% training data)
- Outperformal specialized models like CRAM (96.6%) and SATMAE (96.6%)
- Pixel Timeseries:
- Cropharvest (Kenya): 84.2% (Presto tops and anysat)
- Breizhcrops: 73.0%
- Segmentation (Miou):
- Mados: 67.6%
- Pastis: 79.4%
Flexibility of the model:
In all benchmarks, Galileo is the most efficient overall – by browsing both specialized competitors specializing in image and chronological series. In particular, small model variants (Vit-Nano, Vit-Tiny) also obtain higher or quasi-top results, essential for parameters related to resources.


Ablation and importance of entry
The abolition of any individual modality (for example, VIIRS Night Lights, ERA5, Dynamic World Maps) of pre-training leads to a measurable drop in performance, even on landmarks that do not directly use this type of entry. For example, the absence of VIIRS data reduces Mados Miou by 67.8% to 63.5%, demonstrating the value of the complete multimodality for the generalization of functionalities.
Open-source and real impact
- Open weight and code:
All codes, model weight and pre-training data are available on GithubPromote transparency and adoption by the World OE community. - Societal advantages:
Galileo supports the critical harvesting activities of NASA, such as global mapping of the types of cultures, rapid mapping of disasters (floods, forest fires) and detection of marine pollution. The capacity of the model to work with limited labeled data makes them particularly useful in regions where the truth on the ground is rare, supporting food security and climate adaptation efforts.
Technical summary table
Model | Parameters | Devoted tasks | Rang (lower = better) | Entry procedures |
---|---|---|---|---|
Galileo-Base | 85m | Images, chronological series | 1 (overall) | Optics, SAR, Time, etc. |
Sota specialist | varied | Usually 1 or 2 types | 3–10 | Limit |
Galileo-Base: performance and flexibility always higher than all the main EO references.
Conclusion
The methodological and engineering advances of Galileo – multimodal entries, a large -scale local -globale learning and on a large scale of pre -training on a global scale – establish a new standard for IA general remote sensing. Its flexibility underlies the practical deployments of environmental surveillance to climate resilience, offering reliable and high quality cards and predictions, regardless of the task or geography.
With open source access and active development, Galileo is positioned to catalyze a new wave of innovation in the science of the terrestrial system, which allows practitioners around the world.
Discover the Paper,, Model And Technical blog. Do not hesitate to consult our GitHub page for tutorials, codes and notebooks. Also, don't hesitate to follow us Twitter And don't forget to join our Subseubdredit 100k + ml and subscribe to Our newsletter.
Asif Razzaq is the CEO of Marktechpost Media Inc .. as a visionary entrepreneur and engineer, AIF undertakes to exploit the potential of artificial intelligence for social good. His most recent company is the launch of an artificial intelligence media platform, Marktechpost, which stands out from its in-depth coverage of automatic learning and in-depth learning news which are both technically solid and easily understandable by a large audience. The platform has more than 2 million monthly views, illustrating its popularity with the public.