Image recognition onboard satellites
What are the opportunities and challenges? We explore this across three posts
Thanks for reading Earthbound! If you’re enjoying these posts and want more articles that explore and summarize one paper (instead of a shorter summary of three), consider joining our paid community. And if not, you will still get updates on the latest EO research happening around the world!
This week, we look at a paper that explores the biggest opportunities and challenges that result from onboard image recognition systems. I thought this paper could have provided more detail on the challenges. It goes into great detail on types of AI models, satellite features, and the different components of both, but the challenges related to the main topic were lacking in my opinion (give it a read and let me know if you agree?). Nevertheless, here’s part one of the summary.
First, let’s explore what an on-board image recognition system is. We know that in order to work with earth observation data, we need to download it from some source (ex. NASA Earthdata), then process it, build a model that provides some insight, and finally, have results that we can use. That’s what the EO workflow currently looks like.
What if we didn’t have to do any of that? What if we could “upload” our proven model to the satellite carrying the sensor of interest, and have it return results to us as they’re being collected? With little to no lag time?
This paper describes on-board image recognition systems as an AI model that will “collect, pre-process, and analyze the data directly on the satellite” that is still in orbit and will only return “relevant post-processed data.”
Naturally, with such an innovation, there are challenges. The paper is broken down into three parts, and I’ll break down each into one post.