This study aims to address critical challenges in open-environment synthetic aperture radar (SAR) applications, where high-value targets continuously emerge but labeled samples remain scarce. The study proposes an orthogonality-constrained distribution-calibrated replay method for continuous learning in SAR target recognition. Unlike traditional deep learning models, which assume closed environments with abundant data, this method targets the stability–plasticity dilemma, aiming to resolve the dual problems of catastrophic forgetting of old knowledge and overfitting on limited new data.

First, during the initial base class training phase, the proposed method establishes a highly structured embedding space to reserve non-overlapping geometric regions for future classes. It optimizes the feature space by utilizing a combined orthogonal prototype distribution and intra-class compactness loss. Specifically, the method employs an iterative process to construct an orthogonal basis space by selecting prototype vectors with the lowest cosine similarity to existing bases and projecting them into the orthogonal complement space. This approach ensures that the embedding space is inter-class separable and intra-class compact, leaving adequate room for future incremental classes. When new classes are accommodated, a dual knowledge preservation method is proposed to mitigate catastrophic forgetting of old knowledge during incremental updates by integrating orthogonal gradient projection and distribution-calibrated replay. Orthogonal gradient projection minimizes interference with established knowledge by constraining the update direction. It projects gradient updates into the orthogonal space of the base class feature matrix, such that the model's response to old classes remains invariant. However, achieving a strictly orthogonal space is challenging, and gradual drift still occurs. To further recall old class knowledge from the feature distribution of previous classes, we employ distribution-calibrated replay to reinforce historical knowledge by generating and rehearsing high-fidelity pseudo-features. This strategy utilizes a Gaussian sampler based on the statistics of old classes and introduces a learnable recursive calibrator to correct deviations between the sampled Gaussian distribution and the complex real feature distribution. By minimizing the Kullback–Leibler (KL) divergence between the generated and real distributions, this module replays high-quality pseudo-features that effectively consolidate memory. To enhance the model's robust generalization on scarce new classes, a self-supervised progressive learning strategy is proposed that advances from easy to hard tasks, ensuring effective feature mining from limited samples. This includes a rotation-based self-supervised branch optimized with an uncertainty-weighted focal loss, which effectively optimizes the feature extractor. The self-supervised branch encourages the model to learn azimuth-invariant features, thereby mitigating overfitting. In addition, the uncertainty-weighted focal loss dynamically assigns higher weights to hard-to-classify samples based on prediction probabilities, encouraging the model to focus on mining discriminative features.

Experiment results on the moving and stationary target acquisition and recognition, and SAR AIRcraft-1.0 datasets validate the effectiveness of the proposed method. It significantly outperforms state-of-the-art methods, demonstrating superior efficiency and stability by achieving the highest recognition accuracy with minimal accuracy drop across various scenarios, including different numbers of incremental classes, varying incremental sample sizes, and distinct datasets. Ablation studies confirm that the proposed stability and plasticity modules enhance target recognition performance.

By integrating orthogonal gradient projection, distribution calibration replay, and self-supervised progressive learning, this paper successfully addresses the inherent challenges in continuous learning, providing a robust solution for dynamic SAR automatic target recognition.