The detection of greenhouse gas plumes is a key task of carbon emission monitoring. In the current relevant algorithms, the segmentation accuracy of plume edge details still has room for improvement. In this paper, a multi-scale attention mechanism based plume detection network (MAPDNet) is proposed. Firstly, a multi-scale pooled strip Convolution module (MPSCM) is proposed to compensate the high-level semantic information diluted in the downsampling operation of the encoder, enhance the main plume profile information, and reduce the loss of detail information. Then, the relational attention module (RAM) establishes the global and local information interaction between features, enhances the useful features, reduces the proportion of redundant information in the feature map, and improves the segmentation accuracy of the model. In addition, a context feature fusion module (CFFM) is proposed in this paper so that the network can understand the context feature information better, align the features better, and recover the spatial position information while recovering the image resolution. Finally, the edge enhancement module (BEM) implements end-to-end training, enhances the details of the output segmentation boundary, and further improves the segmentation accuracy. The experimental results show that MAPDNet model can detect carbon plumes from XCO2 plume images, and has good segmentation performance on simulated plume data sets, and its segmentation accuracy is better than the existing methods.