Electroencephalography (EEG) is the most commonly used brain imaging technique for Brain-Computer Interfaces (BCI). Despite its advantages (e.g., safety, cost, time resolution), the intra-subject signal variability poses a major challenge for robustness and accuracy of BCI systems. Two main causes of this variability are the low spatial resolution and the active reference electrode, both responsible for including components from non-task-related cortical areas in the recorded data. In this work, we propose an inverted perspective towards the reference electrode, which relies on looking at a multichannel monopolar set as a collection of information from the reference. Moreover, we propose a heuristic metric that accounts for signal variability and class separability to quantify the potential of each channel to be used for classification between two motor imagery (MI) tasks. The strategy was tested in two datasets and the results reveal strong dependency on the frequency, rather than the scalp region, with the best results occurring for the delta band. Also, the interval which contains the best features for classification depends more on the delay after the MI instruction (~3 to 4 s) than on the activity being performed.