The reason Russia has so much expertise in space (they flew several space stations) stems from the cold war era. During that time, the Russians had Salyut space stations manned with cosmonauts where cameras would fill half of the space station. The reason they had to man the cameras is that the earth is always covered with clouds and the Russians did not have much computing power to process away images with clouds. I talked about this in the following entry: Competing with the human element
The point is that we generally look at clouds as an impediment to imaging. With compressive sensing however, I am wondering the following: Could we use clouds as part of an imaging process ?
The light from the sun bounces from the ground onto the cloud and then some of it bounces back to the ground (cloud albedo).
Could we use some means of determining all the features of these clouds and then use that information to evaluate the cloud albedo. Then use the cloud albedo information to determine the PSF of that system from the sunlight.
The idea comes from the very intriguing paper: The Design of Compressive Sensing Filter by Lianlin Li, Wenji Zhang, Yin Xiang, Fang Li where one could use the different layers of the ionosphere to produce compressed measurements.
I know it sounds a litte crazy but I need to convince myself that it cannot be done by looking at the cloud radiation albedo numbers, so I need to digg this a little further.
3 comments:
I imagine that, if possible, it would be much easier to realize such a PSF estimation with the different layers of the atmosphere (ionosphere) than with clouds.
The reason is that I guess some models of these layers (e.g. their interfaces reflectivity/refractivity) may be assumed sufficiently static, inducing a static PSF (wrt time) in certain EM sensing process.
The fact is that if you use something like "in situ" measurements from some physical medium you do not control, you need a lot of calibration (see e.g. the "random lens imaging" MIT by Fergus and Torralba).
Remember that in CS for instance, even if the matrix is randomly generated, you need to reproduce it latter (e.g. by storing it).
In your framework, if the calibration the "cloud" PSF is slower than the evolution of the cloud configuration, it could be really difficult "deconvolve" the result. This is therefore an interesting problem ;-)
Perhaps some basis facts already exist in the literature, as for using multiple telescopes to reduce atmospheric noise.
My three cents.
Hello Laurent,
Yes, the cloud psf needs to be determined exactly and therefore the "calibration" needs to happen very fast.
In a lidar mode, each measurement goes at the speed of light so I would assume it goes faster than most turbulence of interest in the cloud.
There is a somewhat a related concept in astronomy:
http://www.iop.org/EJ/article/0004-637X/651/1/544/20121.web.pdf?request-id=4316d87c-f3fe-4ffa-9c0f-6d2c44e3bdf8
and more , I'll write on this later.
Cheers,
Igor.
Laurent,
For instance, you can take a look at earthshine studies on the Moon.
http://www.obs-hp.fr/~larnold/publi_to_download/Arnold_2008_SpaceScieRev.pdf
There the Moon acts as a mirror. Can we do better for the "Moon psf" given that we have probes over there ?
Igor.
Post a Comment