<p>In this paper, we propose a novel split federated learning scheme for accelerating the artificial intelligence (AI) model training in low Earth orbit (LEO) satellite networks, named Split-LEO. Specifically, the proposed scheme splits the entire AI model into multiple satellite-side models deployed at LEO satellites and multiple ground-side models deployed at the ground station. Each satellite parallelly performs model training via the collaboration with its corresponding ground-side model and then aggregates satellite-side and ground-side models into a global model, thereby significantly reducing training delay. Furthermore, we formulate an optimization problem with the objective of minimizing training delay via optimizing split point selection and computing resource allocation of satellites and the ground station. To solve the non-convex optimization problem, we transform the problem and then decompose it into two subproblems. The former split point selection subproblem is solved by using an exhaustive search method, while the latter computing resource allocation subproblem is solved by a Lagrange multiplier update algorithm. Extensive simulation results demonstrate that the proposed scheme can significantly reduce training delay while preserving model accuracy as compared with the state-of-the-art benchmarks.</p>