Currently thousands of earth observation satellites are sending data to the surface around the clock. My goal is to make the most out of that data. This story is about one of my recent projects.

In the following example case, the goal was to produce a cloudless view of a specific geographical area using a time series of satellite images. To accomplish this, a sequential approach was used, in which the view is updated step-by-step by utilizing deep learning.

The first image from July 2018 looked like this:

Okay... We can do better than this, I thought. So I took a look at the next image from August 2018...

...and applied the deep pixel fusion algorithm to the first two images:

Much better! I used about 11% of the first image and 89% of the second image. Then, I repeated the fusion process by using the next image in the time series:

old
new
fusion

Once more...

old
new
fusion

...done. The processing was completed with relatively good results. Thanks for reading, until next time! If you want to experience the results of deep pixel fusion in practice, please sign up at terramonitor.com