Parallax measurements of Jupiter's cloud tops in JunoCam images, and applications

JunoCam, Juno's wide-angle visible light camera, has been able to take series of close-up RGB images of the same Jupiter cloud-tops from different angles within only several minutes.
Within this kind of long-baseline observations, cloud motion usually takes a less prominent role than parallax.  
We select features in which it appears reasonable to assume that all relative cloud displacements can be attributed to parallax.  
Although we do not assume that we can determine absolute camera pointing with sufficient accuracy, we can determine relative camera pointing very well, at least locally.  
We first reproject two suitable JunoCam images to the same perspective. Then we stereo-correspond a pair of nearby patches of the first selected image with that of the second image. 
The change of the distances between the two patches returns our desired parallax.
Such stereo-corresponding patches can only be determined in a sufficiently reliable way, if both patches have sufficient small-scale contrast. Cloud-top patches of similar parallax may have an irregular shape. 
We have to deal with these challenges in order to retrieve a sufficiently dense mesh of parallax measurements. 
Because the set of parallax measurements is likely to be noisy and inconsistent, we feed our measurements into an embedded-springs model in order to find a good fit. Spring embedding is more generally known as force-directed graph drawing [1]. In more detail, a parallax measurement relates the elevations of two respective cloud-top patches. That way, we get interconnected islands of relative elevations. The resulting system of equations is likely to be overdetermined and not fully consistent. The spring embedding provides us with the elasticity to retrieve a reasonable solution anyway.
Separating parallax measurements from best-fit calcuations provides the flexibility to refine both portions independently.

A second source of cloud-top altitude estimates comes from JunoCam's methane-band images near a wavelength of 890 nanometers.  
These images are usually noisy and crowded with camera artefacts, most of which are either of a systematic nature or of a point-noise type. In addition there is also statistical photon noise. 
We derive reduced and lower-resolution versions of those methane-band images with most of the camera artefacts removed and some of the photon noise smoothed.  
Those reduced methane-band images can be cross-calibrated with our parallax measurements, whenever we have sufficient overlaps.  
Methane-band images can then be used to fill in cloud-altitude data at a higher resolution where parallax measurements are coarse. 
However, those cross-calibrations can only locally be assumed to be valid since changing observational conditions change the appearence of cloud-top features in methane-band images.
Despite their lack of a photometric calibration, the 890-nm band images provide a qualitative verification of our parallax measurements.

The talk will focus on parallax measurements alone, but we note that observations of small-scale shadows, shading, and hazes along the limb in JunoCam's RGB images also contribute information about cloud-top topography. 
A future extension of the parallax method to JIRAM data appears possible.

[1] Kobourov, Stephen G. (2012), Spring Embedders and Force-Directed Graph Drawing Algorithms, arXiv:1201.3011, Bibcode:2012arXiv1201.3011K