The Allen Institute for AI is using artificial intelligence to map renewable energy infrastructure around the world.

The non-profit research institute has announced Satlas, a platform for exploring global geospatial data generated by AI from satellite imagery. It includes three data products that are updated monthly: Marine Infrastructure positions offshore wind turbines and platforms; Renewable Energy Infrastructure positions solar farms and onshore wind turbines; and Tree Cover gives a tree canopy coverage percentage of each 100m² cell.

Timely geospatial data can be critical for informing decisions in emissions reduction, disaster relief, urban planning and more. For example, renewable energy infrastructure is rapidly expanding across the planet, and accurately tracking this growth across political boundaries is important for prioritising resources and funds.

High-quality global geospatial data products can be hard to find, however. Manually curating these products involves tediously aggregating, cleaning and correcting regional datasets from dozens of countries, assuming these datasets even exist. Indeed, existing geospatial data on renewable energy infrastructure are fragmented, and are only up-to-date in limited geographies.

An alternative is to analyse satellite images from sources such as Landsat and Sentinel-2, which capture images covering most of the Earth every few days. Manual analysis is infeasible due to the sheer volume of data. Automatic analysis of these images has long been error-prone due to their low resolution, where a single pixel corresponds to a 100m² plot of land.

However, given enough training examples, modern deep learning methods can extract data such as the positions of wind turbines from satellite imagery as accurately as humans can.

“Thus, to power Satlas, we have developed a high-accuracy deep learning model for each of the geospatial data products,” said Favyen Bastani, a research scientist at the Allen Institute for AI. “Every month, Satlas applies these models on Sentinel-2 images to derive a new, up-to-date global snapshot of each geospatial data product.”

The institute anticipates that the geospatial data products in Satlas will be useful for a wide range of planetary and environmental monitoring applications. Skylight at AI2 is already exploring whether the marine infrastructure data can be used to improve the classification of vessel movement trajectories, and the institute is looking for other use cases.

Developing high-accuracy deep learning models depends on a large number of high-quality training examples.

“We have manually labelled 36,000 wind turbines, 4000 solar farms, 7000 offshore platforms, and 3000 tree cover canopy percentages in Sentinel-2 imagery,” said Bastani. “We have openly released these training examples, along with the model weights that were learned from them.”

To increase accuracy, the institute leveraged foundation models for Sentinel-2 that it pre-trained on a large-scale remote sensing dataset, SatlasPretrain (satlas-pretrain.allen.ai), which combines several terabytes of Sentinel-2 images with 302 million labels. The foundation models are trained to perform over a hundred tasks simultaneously. These include land cover segmentation, crop type classification and building detection.

The diversity of the images and tasks in SatlasPretrain enables these foundation models to learn descriptive representations of Sentinel-2 satellite images that are robust over different seasons and geographies.

“This means that, when we take a foundation model and train it to perform one particular task very well like detecting solar farms, it offers better and more consistent performance than another model that was trained from scratch,” said Bastani.

The institute has released its Sentinel-2 foundation models along with the SatlasPretrain dataset on which they were trained. SatlasPretrain will appear at the International Conference on Computer Vision (iccv2023.thecvf.com) in October 2023 in Paris, France.

“We have also begun exploring how to enhance the detail of low-resolution but frequent satellite imagery from Sentinel-2, which we call super-resolution,” said Bastani. “In multi-frame super-resolution, we use deep learning models to generate a high-resolution image from many low-resolution images of the same location captured at different times. The model tries to combine information across low-resolution images to predict sub-pixel details.”

It has computed output images from its current super-resolution model globally, and these can be viewed at satlas.allen.ai/superres.

“We plan to continue exploring methods for improving and quantifying accuracy,” said Bastani. “Our primary goal in the short term is to add more geospatial data products to Satlas. We are currently exploring models for mapping urban land use, crop types and land cover, and hope to incorporate a subset of these into Satlas by the end of 2023. We’re also continuing to work on improving the accuracy of the existing data. In the long term, we plan to release tools that make it easier for other teams to build similar geospatial data products, including annotating examples, training models and deploying them.”