The conventional federated learning architecture requires a central parameter server for model collection, aggregation, and distribution. This centralized architecture relies too heavily on a single central server, leading to issues of untrusted central parameter servers and communication congestion. To overcome these challenges, this paper proposed a communication-efficient decentralized hierarchical federated learning architecture, which integrates the blockchain-based decentralized federated learning architecture and the hierarchical federated learning based on edge-end-cloud three-layer architecture. The entire federated learning system is divided into different groups, which are isolated and executed in parallel. The transaction processing, data storage, and block consensus between groups are isolated from each other, and transactions between different groups can be executed in parallel, which ensures the privacy of the blockchain system and improves efficiency. The message exchange between groups will be accompanied by verification information, which is trustworthy and traceable. In addition, in order to improve the communication efficiency of message transmission between different groups, this paper introduced the RAID10 concept in redundant disk arrays, which adopts a strategy of model partitioning and parallel transmission. Divide the clients within the group into multiple communication groups and split the model into multiple parts. The communication groups transmitted a portion of the messages in parallel, further improving the concurrency and transmission efficiency of communication. The relevant experimental platform was built on the Fisco platform, and experiments were carried out on the MNIST data set. The experimental results show that the accuracy rates of 90.63 %, 90.67 % and 91.58 % are achieved respectively in the case of sampling 2, 3 and 5 data samples, and the communication efficiency of message transmission between different groups is also improved.