Decoding　motor　imagery　(MI)　using　deep　learning　in　cortical　level　has　potential　in　brain　computer　interface　based　intelligent　rehabilitation.　However,　a　mass　of　dipoles　is　inconvenient　to　extract　the　personalized　features　and　requires　a　more　complex　neural　network.　In　consideration　of　the　structural　and　functional　similarity　of　the　neurons　in　a　neuroanatomical　region,　i.e.,　a　region　of　interest　(ROI),　we　propose　that　the　comprehensive　performance　of　each　ROI　may　be　reflected　by　a　specific　representative　dipole　(RD),　and　the　time-frequency　spectrums　of　all　RDs　are　applied　simultaneously　to　Random　Forest　algorithm　to　give　a　quantitative　metric　of　each　ROI　importance　(RI).　Then,　the　more　divided　sub-band　spectral　powers　are　reinforced　by　RI,　and　they　are　interpolated　to　a　2-dimensional　(2D)　plane　transformed　from　3D　space　of　all　RDs,　yielding　an　ensemble　representation　of　RD　feature　image　sequences　(ERDFIS).　Furthermore,　a　lightweight　network,　including　2D　separable　convolution　and　gated　recurrent　unit　(2DSCG),　is　developed　to　extract　and　classify　the　frequency-spatial　and　temporal　features　from　ERDFIS,　forming　a　novel　MI　decoding　method　in　cortical　level　(called　ERDFIS-2DSCG).　Based　on　two　public　datasets,　the　decoding　accuracies　of　ten-fold　cross-validation　are　89.89%　and　94.35%,　respectively.　The　results　suggest　that　RD　can　embody　the　overall　property　of　ROI　in　time-frequency-space　domains,　and　ROI　importance　is　helpful　to　highlight　the　subject-based　characteristics　of　MI-EEG.　Meanwhile,　2DSCG　is　matched　well　with　ERDFIS,　jointly　improving　the　decoding　performance.