Spatio-temporal graph learning, a core methodology for addressing urban computing tasks, is often constrained by data scarcity due to factors such as urban development levels or competition. To address this challenge, data from data-rich cities must be leveraged effectively, and knowledge transfer methods should be employed to enhance model performance in data-scarce cities. However, significant differences in spatial features across cities limit the effectiveness of cross-city knowledge transfer. This study proposes the Structure-Aware Enhanced Meta Learning (SAML) framework, a spatio-temporal graph few-shot learning approach. SAML improves model performance in data-scarce target cities through cross-city knowledge transfer. Innovatively, SAML introduces a graph structure-aware enhancement module during meta-learning, addressing spatio-temporal graph structure bias during knowledge transfer by defining a graph contrastive loss. Systematic experiments on three traffic speed prediction datasets demonstrate that SAML outperforms existing state-of-the-art (SOTA) methods.