Ethereum, one of the most widely used blockchain platforms, has become a frequent target for phishing attacks, presenting unique challenges in detecting fraudulent activities in Ethereum transactions. These challenges stem from the complexity of capturing nonlinear temporal patterns in transaction behaviors and mining the intricate global associations between accounts. To address the underlying complexities, we propose the Spatiotemporal Contrastive Framework (SCF), which organizes transactions into time-window sequences and introduces a novel model to capture short-term and long-term temporal patterns. Moreover, through self-supervised learning and multi-view contrastive representation learning, SCF effectively uncovers hidden global relationships among accounts. Experimental results on real-world Ethereum phishing datasets demonstrate that our method achieves nearly a 30% improvement in F1-score and a 70% reduction in false positive rate (FPR) compared to existing approaches.