Multi-branch neural networks are widely used in remote sensing intelligent interpretation because they can fuse multi-modal remote sensing data to improve interpretation accuracy. Meanwhile, in the current “big data era” of remote sensing, the application of distributed deep learning has been steadily growing to improve the training efficiency of remote sensing interpretation models. However, prevailing distributed training systems necessitate users to manually devise a parallelization plan or automatically generate one from a restricted space of model parallelism configurations. At the same time, these techniques are not suitable for extending the application of multi-branch neural network-based remote sensing interpretation on computing devices. In this paper, considering both the data and the model parallelism, we design and implement a fast and near-optimal device placement method for automatically placing branches in multi-branch neural networks across multiple devices. The core idea is to consider the trade-off between the computation and the inter-device communication to minimize the total training time. Through careful formulation of the problem as an integer program, our method can offer optimal placement and scheduling of the branches. We conduct experiments to deploy and evaluate the proposed method on two multi-branch networks. The results indicate that the method finds non-trivial device placements outperforming baselines.