Sequoia has an InvalidPixels tag to identify pixels that behave abnormally (hot, cold, dead, stuck). Although these represent a small number of the total number of pixels.
Any sensor has some inherent random error in its measurement, which is often normally distributed. This directly means that with respect to the mean value of the pixels corresponding to a reflectance target, some will be above and some below the correct value.
Photogrammetry will strongly mitigate this type of error. Since many points of view will contribute many reflectance measurements that are weighed to give a more accurate estimation for a given imaged area. The more samples are collected for a given area, the closer it will be to the mean value.
The usual way of dealing with that on a single-image basis is to simply drop the non-physical pixels. That is, drop any pixels that give reflectances outside of the interval ]0, 1[.
If you have values which are consistently and significantly over 1 then you might be dealing with a non-random measurement error. This could be introduced by an incorrect reference reflectance value, for example.
For any model, this question comes up at some point.
You can expect the mean variation of
B in temperature to be around 200 count over 30K (note that there is a wider variance in the black level than that).
For an adequately exposed image acquired within 20K of working temperature this will represent less than 0.1% error due to temperature in the black level. Parrot considers this to be an acceptable error.