Causal Disentangled Representation Learning for SAR Target Classification With Spurious Correlations

Synthetic aperture radar target classification faces critical challenges from spurious correlations. These erroneous statistical dependencies are established between non-essential features (e.g., background clutter, observation angles) and target labels. Through motivation experiments comparing region-wise contributions to model performance, we reveal that CNN-extracted features exhibit causal entanglement, where background regions serve as sources of spurious correlations. These non-causal features are sensitive to observation angle and background variations under biased conditions, leading to severe degradation in generalization. To address this, causal features should be prioritized for decision-making, and we propose the Causal Disentangled Representation Learning framework (CausalDRL) for biased SAR classification. CausalDRL introduces a structural causal model to formalize the biased SAR classification task, based on this, a learnable binary mask disentangler is proposed to decouple features into causal and bias subsets. Then, to maximize the decoupling degree, a shuffle strategy is utilized to generate counterfactual features, while a batch-weighted average label mechanism is proposed to suppress label-related information in bias features, thereby further purifying the disentanglement. For controlled validation, the first SAR dataset with adjustable background bias degrees, SARBake-bias, is proposed. Comprehensive evaluations on SARBake-bias ($\sim$10%$\uparrow$), ${\rm{SARBake-bias}}_{\rm{50}}$ ($\sim$20%$\uparrow$), ATRNet-STAR, MSTAR, OpenSARShip, FUSARShip, and SARAircraft-1.0 demonstrate the state-of-the-art performance. The framework provides interpretable spurious correlation mitigation without additional prior knowledge dependency, significantly advancing biased SAR classificationn. The data, code and weight files will be gradually released at https://github.com/Ivan12138a/CausalDRL.