As blockchain technology advances, the use of smart contracts has increased dramatically across many different industries, and Ethereum has become the most popular smart contract platform. However, frequent smart contract vulnerabilities have caused huge economic losses, so it is necessary to study effective smart contract vulnerability detection technology. Existing methods cannot fully utilize the grammatical information of the contract source code to detect smart contract vulnerabilities. This study introduces SCDefender, a deep learning-based vulnerability detection tool that solves the smart contract re-entrancy problem. SCDefender uses spatial block and spatial correlation learning to improve its vulnerability detection efficiency. It uses tree-based convolutional neural networks to find bugs in the abstract syntax tree form of Solidity source code. In order to eliminate redundant nodes and retain important data in the abstract syntax tree, pruning technology is also introduced. SCDefender is a promising solution for identifying smart contract vulnerabilities because it combines spatial associative learning and spatial blockchain. SCDefender performed well, with precision, recall, and F1 scores of 81.43%, 92.12%, and 86.45%, respectively. Furthermore, sensitivity analysis shows that the vulnerability detection performance of SCDefender greatly benefits from the abstract syntax tree pruning approach.