Electroencephalogram (EEG) have gradually become one of the mainstream research hotpots in emotion recognition owing to its high time resolution, reliability, and rich information. However, most studies construct brain networks based on EEG full-channel, often overlooking the high-order interactions between regions, the information flow within regions, and high-dimensional relationships between global and local scales. To tackle this problem, we propose a topology-based region hierarchical convolutional network (TRHCN) for EEG emotion recognition, which utilizes simplicial complex to represent the high-order structure at global and local scales, and employs heterogeneous graph to establish connections for global and local interactions. Specifically, we design three encoding modules to extract discriminative emotional features at different levels, i.e., a global topology encoder module, a local attention encoder module, and a heterogeneous graph convolution module. Experimental results on SEED and SEED-IV datasets indicate that the TRHCN model effectively captures the underlying interactions at three scales related to regions and outperforms other approaches.