14.08.23 - Prithvi: Further Developments in AI-based Analysis of Satellite Images

Have you ever wondered what the Earth would look like in 10, 20, or 50 years? How will climate change affect our planet and our lives? How can we use data and AI to understand and predict the complex dynamics of our environment?

These are some of the questions that IBM and NASA - National Aeronautics and Space Administration are trying to answer with their new project: #Prithvi. Prithvi is a first-of-its-kind temporal Vision transformer pre-trained on a multi-band satellite image dataset, which is capable to deal with additional geospatial layers like NIR and SWIR. They hope this open-source infrastructure will serve as the fundamental for future forest, crop, and climate change-monitoring AI. 

Leveraging NASA's HLS V2 L30 product (30m granularity) from the contiguous United States, Prithvi has access to petabytes of satellite and ground-based observations. The model adopts a self-supervised encoder developed with Masked AutoEncoder (MAE) learning strategy. With the pre-trained ViT encoder, Prithvi can give considerable promising results with little effort on finetuning the exact downstream task. 

Examples of image segmentation applications are available through Hugging Face (e.g. burn scars segmentation, flood mapping, and multi-temporal crop classification).

How do you think this kind of AI can help us better understand and protect our planet? STDL is planning to launch projects exploring the possibility of a similar framework with satellite and aerial images across Europe. If you have any interest, feel free to contact us!

#climatechange #vision-transformer #burn-scars-segmentation #flood-mapping #crop-classification