In the study of model parameter based transfer learning, both the sample distribution discrepancy between two domains and the co-adaptation between convolutional layers of the source model impact performance of model. In response to these problems, a Multi-Convolution Adaptation (MCA) deep transfer framework was proposed and applied to the grade classification of typhoons in satellite cloud images, and a CE-MMD loss function was defined by adding the improved L-MMD (Maximum Mean Discrepancy) algorithm as a regular term to the cross-entropy function and applying the linear unbiased estimation to the distribution of the samples in Reproducing Kernel Hilbert Space (RKHS). In the back propagation process, the residual error and the distribution discrepancy between the samples in two domains were used as common indexes to update the network parameters, making model converge faster and have higher accuracy. Comparison experimental results of L-MMD and two measurement algorithms-Bregman difference and KL (Kullback-Leibler) divergence on the self-built typhoon dataset show that the precision of the proposed algorithm is improved by 11.76 percentage points and 8.05 percentage points respectively compared to those of the other two algorithms. It verifies that L-MMD is superior to other measurement algorithms and the MCA deep transfer framework is feasible.