Myocardial infarction (MI) can cause acute and permanent damage to the myocardial muscle. Vectorcardiogram (VCG) is formed by the time-varying coordinates of cardiac electrical activity in space. According to different infarct locations, the ring of VCG in the three orthogonal planes has pathological morphological changes. Yet the existing algorithms only extract the pathological information of three-lead VCG signals, but they do not fully consider the correlation information between different orthogonal planes. We proposed a depthwise separable convolution network (DSC-Net) for automatic MI localization from 2D-VCG tensor. Using the orthogonality between the lead axes, we first combine the three leads in pairs to form a 2D-VCG, and then construct a 2D-VCG tensor that captures the correlation information between leads. DSC-Net extracts spatial features related to MI obtained in 2D-VCG before Softmax is applied to classify MIs. The proposed method was validated on the benchmark Physikalisch Technische-Bundesanstalt dataset, which includes VCG of 11 types of MI. We demonstrated, with the cardiac electrical activity spatial features obtained from the 2D-VCG tensor, that the accuracy of 11 categories of MI and normal is higher than 99.92%. The proposed model effectively realized the localization of MI with competitively high accuracy for all 11 categories.