Mercator Labs: Affordable geospatial foundation models

About Mercator Labs

Who we are: A tech startup building affordable geospatial AI infrastructure.

What we do: We develop geospatial foundation models that understand satellite imagery and location data—similar to AlphaEarth and Prithvi, but dramatically cheaper.

Problems we solve: Advanced Earth observation AI is expensive and inaccessible. Researchers, startups, and organizations working on climate monitoring, agriculture, disaster response, and urban planning can't afford enterprise-scale geospatial AI.

Target audience: Research institutions, climate tech startups, agricultural technology companies, disaster response organizations, and government agencies needing cost-effective geospatial intelligence.

Our technology:

Satellite Imagery
Multi-spectral data from Landsat, Sentinel, and other sources.
Pre-trained on petabytes of global imagery.

Location Embeddings
Geographic coordinates encoded as dense vectors.
Captures spatial relationships and context.

Self-Supervised Learning
Masked autoencoders for efficient pre-training.
No labeled data required for foundation training.

Fine-tune for your specific task with 10-100x less labeled data than training from scratch.

Mercator Platform

What we're building: An API platform for deploying and fine-tuning geospatial foundation models on custom data.

Current stage: Private Beta

Key features:

Pre-trained Models API: Access foundation models via REST API for inference and embedding generation
Easy Fine-tuning: Upload your custom data and fine-tune models with minimal code
One-Click Deployment: Deploy fine-tuned models as scalable API endpoints
Python SDK & Notebooks: Libraries and examples for rapid prototyping

Tech Stack: JAX, TPUs, optimized for cost-efficient training and inference

Get access: Email pinak@mercatorlabs.xyz to join the private beta and receive API credentials, documentation, and example code.

Team

Pinak Paliwal - Founder
Current student at UC Berkeley MET. Has worked on geospatial foundation models for the past year, focusing on cost-efficient architectures for satellite imagery analysis.

Based in Berkeley, California.

What can you do?

Our foundation models can be adapted for:

Land cover classification and semantic segmentation
Change detection and temporal analysis
Crop monitoring and yield prediction
Disaster response (floods, fires, infrastructure damage)
Urban planning and development tracking
Environmental monitoring and conservation
Geospatial similarity search
Location-based intelligence and forecasting

Higher accuracy than AlphaEarth/OlmoEarth/similar models on geospatial tasks.
Less training data required. Faster convergence.

Examples

GeoGuessr Demo:
We've fine-tuned our foundation model to predict geographic locations from street view imagery. Try it out at geoguessr.mercatorlabs.xyz.

This demonstrates how our base geospatial model can be adapted for location intelligence tasks with minimal fine-tuning data.

Pricing & Access

We're currently in private beta.

Email pinak@mercatorlabs.xyz to get access.

We offer flexible pricing based on your use case:
Research? Commercial deployment? Custom fine-tuning?
Let's talk.

FAQ

What is a geospatial foundation model?: A geospatial foundation model is a large neural network pre-trained on massive amounts of satellite imagery and location data. Similar to how GPT understands language, our models understand Earth observation data. We pre-train on unlabeled data from across the globe, which means you can fine-tune for your specific task with 10-100x less labeled data than training from scratch.
How is this different from AlphaEarth or other models?: We optimize for cost efficiency. Models like AlphaEarth demonstrate incredible capabilities, but require significant compute resources that most researchers and small companies cannot access. We focus on maximizing performance per dollar to make geospatial AI genuinely accessible.
What kind of data do your models work with?: Any georeferenced raster data. This includes multi-spectral satellite imagery (Landsat, Sentinel-2), SAR data, elevation models, and more. Our location embeddings can encode any latitude/longitude coordinate into a dense vector representation.
Do I need labeled data to use your models?: Not for the foundation model itself, which comes pre-trained. For fine-tuning to your specific application, you will need some labeled examples, but typically 10-100x fewer than training from scratch. This is the core value proposition of foundation models.
What tasks can I fine-tune for?: Any geospatial analysis task. Classification, semantic segmentation, object detection, temporal change detection, regression tasks like crop yield prediction. You can also use the embeddings directly for similarity search and clustering applications.
How do I get started?: Email pinak@mercatorlabs.xyz with a description of your use case. We will provide API access or model weights, along with documentation and example notebooks to help you get started quickly.
Can I use this for commercial applications?: Yes. We offer commercial licensing with pricing based on your deployment scale and requirements. Contact us to discuss your specific needs.
Do you offer custom model training?: Yes. If you have specific requirements like custom geographic regions, different spectral bands, or specific temporal resolutions, we can train custom models. Get in touch to discuss.
What about inference speed and deployment?: Our models are optimized for efficient inference on standard GPU infrastructure and major cloud platforms (AWS, GCP, Azure). If you have specific hardware constraints, we can help with optimization.
Is this research-friendly?: Absolutely. We offer academic pricing and actively support research projects. Many critical problems in climate science, agriculture, and disaster response require accessible tools. Enabling that research is part of our mission.
What is your goal?: To democratize geospatial AI. Currently, advanced Earth observation capabilities are limited to organizations with large compute budgets. But researchers working on climate change, sustainable agriculture, and disaster response need these tools too. We are working to close that gap.

Questions? Reach out: pinak@mercatorlabs.xyz