Understanding AI models for satellites and associated data

What is it about?

This is an overview of AI models that train on what we would call "Earth Observation" data. This typically refers to satellites, but realistically refers to anything that is making observations about the planet (such as weather stations). In the past 5+ years, a number of researchers have started training models to learn how to understand earth observation data. Understanding this research requires an understanding of state of the art models in computer vision, a plethora of satellite sensors, and how the two pertain to one another. This paper is designed to give a bit of a crash course in understanding all of the above. It particularly focuses on how existing approaches in computer vision have been adapted to work with satellite imagery.

Featured Image

Photo by Javier Miranda on Unsplash

Photo by Javier Miranda on Unsplash

Why is it important?

There is an enormous amount of publicly available satellite data, to the point that it can become cumbersome to leverage it. For instance: a single satellite generates roughly 1.6 TB of data per day. Working with that scale of data if you're not a computer scientist can be challenging. But the thing is, this data contains an incredible amount of information. It can be used for flood plain mapping, wildfire recovery forecasting, biodiversity monitoring, air pollution monitoring, sea ice detection for shipping, and an enormous amount of other tasks. Models like the ones outlined in this paper make it easier for people to perform these tasks. This paper should be a nice source of reference for people trying to understand what these models are good at. This is also a burgeoning research field (for all of the reasons outlined above). Yet it can be daunting to approach because it requires broad expertise across a variety of fields. It's our hope that this paper serves as a good entry point for researchers who are entering this particular domain.

Perspectives