To　address　the　issues　of　low　spatial　resolution　and　susceptibility　to　noise　in　traditional　single-modality　brain-computer　interface　(BCI)　technologies　based　on　electroencephalography　(EEG),　an　increasing　number　of　studies　have　focused　on　BCI　research　that　combines　EEG　signals　with　functional　near-infrared　spectroscopy　(fNIRS)　signals.　However,　integrating　these　two　heterogeneous　signals　poses　challenges.　This　paper　proposes　an　innovative　end-to-end　signal　fusion　method　based　on　deep　learning　and　evidence　theory　for　motor　imagery　(MI)　classification.　The　spatiotemporal　feature　information　of　EEG　signals　is　extracted　using　dual-scale　temporal　convolution　and　depth　wise　separable　convolution,　with　a　hybrid　attention　module　introduced　to　enhance　the　network’s　ability　to　perceive　important　features.　For　fNIRS　signals,　spatial　convolution　across　all　channels　explores　activation　differences　between　different　brain　regions,　while　parallel　temporal　convolution　and　gated　recurrent　unit　(GRU)　capture　richer　temporal　feature　information.　During　the　decision　fusion　stage,　the　decision　outputs　obtained　from　decoding　each　signal　are　first　utilized　to　estimate　uncertainty　using　Dirichlet　distribution　parameter　estimation.　Subsequently,　Dempster-Shafer　theory　(DST)　is　employed　for　dual-layer　reasoning,　effectively　merging　evidence　from　the　two　basic　belief　assignment　(BBA)　methods　and　different　modalities　to　obtain　the　decoding　results.　The　proposed　model　is　evaluated　on　the　publicly　available　TU-Berlin-A　dataset,　achieving　an　average　accuracy　of　83.26%,　which　represents　a　3.78　percentage　points　improvement　compared　to　the　state-of-the-art　research.　This　provides　new　ideas　and　approaches　for　fusion　studies　based　on　EEG　and　fNIRS　signals.　©　2025　Chinese　Institute　of　Electronics.　All　rights　reserved.