There are several other opinions. Please, evaluate some papers at Attachment. All  the best.

Thank you very much Sir!

With my pleasure!

I would propose to look at some of the original neuroscience literature on imagined movement, e.g. by Pfurtscheller:

https://www.google.nl/search?q=pfurtscheller+motor&ie=utf-8&oe=utf-8&gws_rd=cr&ei=LoP1VLLGH8XyOOSIgIgG

and then at some papers about BCI and motor imagery, e.g. by the Berlin group:

http://doc.ml.tu-berlin.de/publications/klaus-robert.mueller

Short answer: C3 and C4 are the most commonly used sensors. The Berlin group in particular fancies a mathematical approach called common-spatial patterns, which extracts a spatial feature map that is then used for each subject. The purpose of your research, however, defines your approach. If you are an engineer, you might want to try the heavy machine learning approach. If you are a neuroscientist you might want to care more about the underlying brain region and related brain mechanism.

Generally, you will want to use channels that contain useful information about that particular paradigm. For the P300 BCI experiments,  according to my experience, the most relevant channels are Fz, Cz, and (especially) Pz. But other channels may help you to better deal with artifact removal, e.g. EOG for eye motion/blinking removal, or O1 and O2 channels for alpha waves.

That depends largely on the application for placement of electrodes, for example the frontal part of the brain is very used for imagery motor, affective and emotional signals, i.e. passive or active BCIs; unlike of the occipital region mostly used for reactive BCIs .

Thanks all, the comments are very informative and I have learned and applied these suggestions! :)