AI in Satlas
Satlas combines the power of modern AI with the scale of public domain satellite imagery to provide monthly monitoring of the planet.
AI models in Satlas process freely available satellite images captured by the European Space Agency’s Sentinel-2 satellites. These images cover the majority of the planet every week, but are low in resolution, making them difficult to interpret even for humans.
We leverage the latest advances in AI to robustly process images in this challenging domain and produce a variety of geospatial data products that we make freely available.
SatlasPretrain
Our AI models are pre-trained on a new large-scale remote sensing dataset called SatlasPretrain. This vast dataset contains over 30 TB of imagery with 302 million labels spanning 137 diverse categories, from tree cover and crop fields to wind farms and oil wells. Pre-training on SatlasPretrain teaches AI models to understand geographically and seasonally diverse satellite images.
Fine-tuning the Models
The AI models in Satlas are fine tuned on high-quality training datasets that we hand-label for each geospatial data product. In practice, we use multiple steps of fine-tuning, testing, and additional labeling to progressively improve the accuracy of our models.
Access the Training DataAI Model Architecture
Our AI models use state of the art machine learning architectures and training methods. They input a sequence of the three most recent satellite images captured at each location. Each image is passed through a Swin Transformer backbone to extract features. These features are then combined via max temporal pooling, and passed to task-specific neural network heads to make the final predictions.
Data Accuracy
A per-continent breakdown of the accuracy of each Satlas geospatial data product is made available in our Data Validation Report. We also showcase examples of the limitations of our data, including erroneous and missing data. Overall, the data generated by our AI models has high accuracy, but AI systems are never perfect and several factors tend to degrade performance.
View Data Validation Report