At　present,　how　to　efficiently　and　effectively　identify　motor　imagery　tasks　is　still　a　huge　challenge　for　the　development　of　the　Brain　Computer　Interface　(BCI)　systems　in　the　field　of　human　rehabilitation.　Therefore,　this　paper　proposes　an　optimized　recognition　method　based　on　temporal　features　and　spatial　features　re-representation.　The　Local　Mean　Decomposition　(LMD)　algorithm　is　used　to　extract　the　Product　Functions　(PFs)　of　the　Motor　Imagery　Electroencephalogram　(MI-EEG)　signals　and　the　Common　Spatial　Pattern　(CSP)　algorithm　is　applied　to　reconstruct　the　spatial　distribution　of　each　PF,　then　the　MI-EEG　signals　are　re-represented　as　features　with　temporal　and　spatial　characteristics.　The　Probabilistic　Neural　Network　(PNN)　is　constructed,　in　which　the　smoothing　factor　is　optimized　by　the　Particle　Swarm　Optimization　(PSO).　By　introducing　the　PSO　algorithm,　the　PNN　can　be　adaptively　determined　according　to　the　respective　conditions　of　different　subjects　or　datasets.　The　experimental　results　show,　compared　with　the　Support　Vector　Machine　(SVM)　based　on　feature　re-representation　and　other　state-of-the-art　machine　learning　methods,　the　proposed　method　in　this　paper　has　both　high　recognition　accuracy　and　adaptability.　This　optimized　PNN　with　PSO　method　establishes　a　theoretical　foundation　and　methodological　guidance　for　the　decoding　of　motor　imagery　recognition.　©　2023　IEEE.