Nonmuscle　channels　are　able　to　enhance　voluntary　movement　control　or　improve　rehabilitation　efficacy,　and　analyzing　electroencephalogram　(EEG)　signals　yields　valuable　insights　into　functional　neural　activity　during　motor　imagery　(MI).　However,　most　existing　MI-EEG　datasets　primarily　focus　on　exploring　more　discriminative　tasks　to　control　external　devices,　making　it　difficult　to　meet　the　actual　needs　of　motor　function　rehabilitation.　In　addition,　the　expensive　acquisition　of　labeled　data　hinders　the　practical　application　of　EEG-based　rehabilitation.　In　this　study,　we　aim　to　collect　EEG　signals　from　13　subjects　performing　four　MI　tasks　of　the　unilateral　upper　limb:　arm　lifting/lowering　and　forearm　pronation/supination.　Moreover,　we　propose　a　Contrastive　representation　learning　framework　with　an　Attention　Spatiotemporal　Convolutional　Encoder　(CASCE)　for　MI-EEG　decoding.　In　the　pretraining　phase,　the　unlabeled　data　from　background　samples　and　label-erased　samples　are　applied　to　the　noise　addition　module　and　the　scaling　module　to　generate　pairs　of　positive　and　negative　samples.　These　sample　pairs　are　input　into　the　encoder　to　learn　temporal　and　spatial　information,　and　the　encoder　parameters　are　further　adjusted　by　using　the　contrastive　loss　function　to　measure　the　similarity　of　the　feature　information　in　the　projection　space.　During　the　fine-tuning　phase,　the　transferred　encoder　and　the　classification　head　are　specifically　adapted　to　the　labeled　MI-EEG　data.　The　CASCE　framework　achieves　a　classification　accuracy　of　51.58%　on　the　refined　upper　limb　MI　dataset.　In　addition,　CASCE　outperforms　state-of-the-art　(SOTA)　methods　with　accuracies　of　88.51%　and　90.34%　on　the　brain-computer　interface　(BCI)　Competition　IV　2a　and　2b　datasets,　respectively.