From 2016 to 2018
Earth observation satellites generate a lot of data and most analytic solutions are automated. I briefly touched two sensitive areas of research in this field, the cloud detection and land labeling. I had the honor of working with the following instruments/satellites: GOCI (Korea Ocean Satellite Center), MODIS, CALIPSO, Sentinel 1, Sentinel 2, and Planet.
Clouds are formed by liquid water aggregated in a large mass inside the atmosphere. They are tricky to be detected because in this form, water “looks” roughly the same as snow, smog or some marine currents. Not to mention parallax effects and cloud shadows. Of course, some instruments have spectral bands dedicated to this discrimination, but not all.
The goal of the project was to develop a better “cloud mask” for a specific instrument, GOCI, that is on a geostationary satellite looking over the Asia region.
The interesting take about this project was that the progress was done through massive feature engineering rather than ML. Despite the not-so-great success, I got the pleasure of staring at the clouds for my day job!
NASA provides a nice data visualization for GOCI archive. Feast your eyes with the endless and weird shapes the clouds can form!
My favourite? Kármán vortices and early winter sunrises!
Firey sunrise [or sunset] over China(?) during winter. The GOCI instrument aquire images only during the day. However, in winter, the sun is really low, creating these spectacular images! The reds are because our atmosphere scatters the blue wavelengths a bit more.
von Kármán vortices created by Yakushima(?) island.
Unfortunately I lost track of where and when those scenes were originated from and I don’t have the full composite GOCI image archive. I tried to browse NASA archive, but well, searching the needle in a quite large image haystack.
Land mass characterization
An application when having multi-spectral, multi-modality data is to determine the quality of the surveyed land. For example, quality and quantity of crops could be evaluated automatically and government level entities could act on such analytical data.
With my team I tried to do a “sensor fusion” between Sentinel 1 and 2 data. With high quality labels and enough compute power, the results were decent. Looking “through” the clouds, and “seeing” changes is quite a powerful tool. Unfortunately, the stakeholders decided that the derived analytics are not relevant enough [and did not fund the prototype] so the project was cancelled.