Cloudless: Open Source Deep Learning Pipeline for Orbital Satellite Data
I'm proud to announce the 1.0 release of Cloudless, an open source computer vision pipeline for orbital satellite data, powered by data from Planet Labs and using deep learning under the covers. This blog post contains details and a technical report on the project.
Read more...
Read more...
Comments
1) Automatic tie-point detection. If we are able to spatially tie regions of the image to reference imagery, those regions probably don't have clouds.
2) We've generated statistics from historical Landsat 8 imagery for the expected brightness at any given lat/long and day-of-year. So, to generate a cloud mask, we just look up the number of standard deviations from that expected value for every pixel. Yes, I know it doesn't work well over land cover types like snow & ice, but it works pretty good otherwise, was easy to code up, and is very fast. I've been planning on investigating deep learning when I have some spare time.
Each of the above methods are orthogonal to your normal top-shelf machine learning algorithms such as neural nets. So, I'd be curious how your accuracy compares to our cloud masks and if your accuracy improves if you incorporate them as auxiliary data.
Also, I would expect that the 90deg rotated imagery to be lower accuracy because of cloud shadow directions. The neural nets are probably keying in on the fact that the sun is always coming from a southern direction.