Non invasive brain-computer interfaces (BCI) allow people to communicate by modulating features of their electroencephalogram (EEG). Spatiotemporal filtering has a vital role in multi-class, EEG based BCI. In this study, we used a novel combination of Principle Component Analysis, Independent Component Analysis and Dipole Source Localization to design a Spatiotemporal Multiple Source Tuning (SPAMSORT) filter bank, each channel of which was tuned to the activity of an underlying dipole source. Changes in the Event-Related Spectral Perturbation (ERSP) were measured and used to train a linear Support Vector Machine to classify between four classes of motor imagery tasks (left hand, right hand, foot and tongue) for one subject. ERSP values were significantly (p<0.01) different across tasks and better (p<0.01) than conventional spatial filtering methods (Large Laplacian and Common Average Reference). Classification resulted in an average accuracy of 82.5%. This approach could lead to promising BCI applications such as control of a prosthesis with multiple degrees of freedom.