Foundation cloud picture classification method based on attention convolution

The invention discloses a foundation cloud picture classification method based on attention convolution, and the method comprises the steps: selecting a multi-stage classifier as a basic architecture, employing an improved re-parametric deep convolutional neural network as a backbone network, and dividing the method into a training stage and a reasoning stage. In the training stage, the ground-based cloud depth representation convolutional network adopts a multi-branch topological structure with rich gradient information and a complex network structure to extract ground-based cloud picture features; in order to obtain the global feature representation of the image, a new efficient channel attention network is introduced, and the feature relationship between the sequences is learned; besides, a learning rate preheating algorithm is introduced in the training stage of the CloudRVE, the model learning rate can be dynamically optimized, and the method aims at accelerating model parameter convergence and enhancing