Recently, graph neural networks (GNNs) have faced significant challenges in handling out-of-distribution (OOD) generalization problems. Traditional methods often learn spurious correlations due to the confounding effects of unobserved environmental variables. Although causal intervention frameworks (e.g., CaNet) have partially addressed this issue through variational inference and backdoor adjustment, their environmental estimators still rely on graph convolution operations, which struggle to effectively capture multi-scale environmental features. Furthermore, existing approaches lack synergistic modeling of local topological structures and frequency-domain information when disentangling environment-sensitive features, resulting in limited generalization capability. To address these challenges, this paper proposes SEE-Net, a causal intervention model that integrates graph wavelet transform with contrastive learning. Inspired by graph wavelet neural networks, we introduce graph wavelet transform as the core component for environmental feature extraction. Through multi-scale decomposition using wavelet basis functions, the model simultaneously captures local high-frequency details and global low-frequency trends in node neighborhoods. This joint spatial-frequency representation significantly enhances the environmental estimator’s ability to identify implicit confounding factors. To further improve the acquisition of environmental features, we incorporate contrastive learning into the causal framework. Drawing on the idea of node-label contrastive encoding, we construct a contrastive learning mechanism based on node environment features. Experimental comparisons demonstrate that our model achieves maximum improvements of 2.57% and 2.87% over the baseline models under GCN-backbone and GAT-backbone configurations, respectively.