<p>The probabilistic stability analysis of levees is hindered by high computational cost of finite element methods, especially when considering the spatial heterogeneity of soil properties. While deep learning-based surrogate models can significantly improve computational efficiency, traditional convolutional neural networks-based models often struggle to capture the complex, multi-scale dependencies between random material fields and response fields, particularly under small sample training conditions. To address this problem, a physics-guided image-to-image surrogate modeling framework is proposed. First, an improved Karhunen–Loève expansion approach is developed by integrating statistical moment matching and matrix decomposition-based cross-correlation embedding, enabling the efficient generation of non-Gaussian and cross-correlated random fields. The generated random fields are mapped onto finite element meshes and reconstructed into grayscale images, ensuring spatial topological consistency while enabling efficient digital input for the surrogate model. Subsequently, an enhanced U²-Net surrogate model is established to learn the nonlinear mapping relationship between heterogeneous material parameter fields and corresponding response fields. The results demonstrate that the proposed framework serves as an effective predictor for seepage and stress fields and can accurately identify high risk regions in the levee that act as precursors to failure under small-sample conditions. This work provides a low-cost and high-precision framework for the risk assessment of hydraulic infrastructure.</p>