GeoFMs on AWS Redefine Scalable Earth Observation Intelligence

By Chris Erasmus, Country General Manager, AWS United Arab Emirates & RoMENA 

For years, Earth observation workflows depended on building task-specific models—an approach that was both resource-intensive and difficult to scale. Today, transformer-based vision models are reshaping how we monitor the planet.

Geospatial foundation models (GeoFMs)—such as Clay, Prithvi-100M, SatMAE, AlphaEarth, OlmoEarth, and SatVision-Base—are driving this shift through self-supervised learning. By pre-training on vast volumes of unlabeled data, they learn the underlying patterns, textures, and spatial relationships within geospatial imagery. The outcome is a new class of models that inherently understand what the Earth looks like and can be fine-tuned for targeted use cases with significantly less data and time than traditional approaches.

Amazon Web Services (AWS) provides the specialized infrastructure necessary to handle the unique demands of GeoFMs. These transformer-based vision models offer a new way to map the earth’s surface at continental scale.

The Shift to Foundation Models

Historically, analyzing satellite imagery required supervised learning, where experts manually labeled thousands of images to teach a model to identify specific features. This approach is often brittle, as models trained on one geographic area frequently fail when applied to another.

GeoFMs leverage masked autoencoders (MAE) to pre-train on unlabeled geospatial data sampled globally. This self-supervised approach ensures diverse ecosystems and surface types are represented, creating general-purpose models that understand Earth’s fundamental patterns without requiring extensive labeled datasets for every new application.

Scaling Earth Observation with AWS

AWS is designed to provide specialized infrastructure to handle the unique demands of GeoFMs, which involve massive file sizes and complex coordinate systems. Data at Scale: Through the Registry of Open Data on AWS, users access petabytes of imagery (like Sentinel-2) without moving it. This “data-gravity” approach minimizes latency and egress costs. Purpose-Built Tooling: Amazon SageMaker offers integrated environments to build, train, and deploy these models. SageMaker AI Pipelines supports the automated “chipping” of raw imagery into manageable 256×256 pixel segments for analysis. Compute Power: Training GeoFMs requires intense GPU resources. AWS GPU instances are designed to provide distributed computing capabilities to process global-scale datasets efficiently.

Core Use Cases for Planetary Intelligence

The integration of GeoFMs on AWS supports three core capabilities:

• Geospatial Similarity Search: GeoFMs convert imagery into high-dimensional vector embeddings. This allows for “image-to-image” searching where a user can select a reference area—such as a specific crop type or an area of urban sprawl—and instantly find similar patterns across vast territories.
• Embedding-Based Change Detection: By analyzing a time series of embeddings for a specific region, analysts can pinpoint exactly when and where surface disruptions occur, such as identifying early signs of forest degradation before they expand into large-scale clearing.
• Custom Machine Learning: Organizations can fine-tune a lightweight “head” on top of the GeoFMs. This allows for high-accuracy tasks like semantic segmentation (classifying every pixel in an image) with significantly less training data than traditional models.

Real-World Impact

The practical application of these models is already driving innovation. In the Amazon rainforest, researchers are using the Clay foundation model on AWS to detect subtle signatures of selective logging and new access roads. This early detection allows environmental protection agencies to deploy resources precisely to prevent major forest loss.

The solution is highly adaptable; while current examples focus on the Amazon, the same pipeline architecture works seamlessly with various satellite providers and resolutions to address challenges across industries like agriculture, insurance, energy and utilities, disaster response, and urban planning.

The Future of Earth Observation

While geospatial data pipelines remain essential, GeoFMs on AWS dramatically reduce the burden through shorter training cycles with fine-tuning or zero-training approaches like embedding-based similarity search. This enables organizations to focus on solving pressing environmental and economic challenges. The technology is ready. The question now is how quickly organizations will adopt these tools to address these challenges that demand immediate action.