Synthetic Aperture Radar target classification faces dual challenges under dynamic and complex environments: environment interference and cross-domain feature drift. Existing methods, relying on statistical correlations, often fail to disentangle spurious environmental features from target-invariant causal features, leading to limited the robustness of models. This paper proposes an Environment-Invariant Causal Feature Extractor (Env-CFE) to address these issues through causal inference. In Env-CFE, a novel structural causal model is first constructed to formalize the problem, and based on this, a causal mask is generated to disentangle the original features into causal and redundant subsets. Furthermore, counterfactual features are designed to large the distance between these two subsets, and a purified label is introduced to further squeeze the non-target attribute features from the extracted latent causal features. Extensive experiments on the MSTAR, OpenSARShip, SARAircraft-1.0, and SARBake-Env datasets demonstrate the framework’s robustness against environmental shifts, outperforming state-of-the-art methods in cross-domain generalization. The results validate the effectiveness of extracting latent causal features to suppress shortcut learning and improve recognition reliability in real-world scenarios.