Smart contracts are self-executing programs that run on blockchain platforms. As smart contracts are increasingly applied across various fields such as finance and agriculture, concerns regarding their security have become more prominent. Leveraging deep learning for detecting vulnerabilities in smart contracts has gradually emerged as a trend, particularly through the use of Graph Convolutional Networks (GCN). However, the current application of graph convolutional networks in smart contract vulnerability detection predominantly focuses on node-level aggregation, which often leads to suboptimal detection performance. While incorporating expert knowledge into the feature extraction process of GCN can enhance accuracy, the detection performance and scope are partially dependent on this expertise, making it less suitable for comprehensive vulnerability detection. To address these challenges, this paper proposes a detection method for Ethereum smart contract vulnerabilities. Specifically, we first construct a Directed Contract Graph (DCG) based on the Abstract Syntax Tree (AST). Then, leveraging the structural characteristics of DCG, we introduce an Augmented Graph Convolutional Network (AugGcn) and Degree-Sensitive Attention Pooling (DSAP) to extract and aggregate rich structural and semantic features. Finally, classification is performed using a Multi-Layer Perceptron (MLP) to obtain the detection results. We conducted a substantial number of experiments. The results indicate that our method outperforms other advanced detection techniques currently available, achieving an average accuracy of 93.8%.