A Novel CycleGAN-Based Domain Adaptation Method for SAR Incremental Learning

Spaceborne SAR ship detection relies on deep learning algorithms, but practical deployment necessitates Incremental Learning (IL) due to continuous introduction of new data domains (e.g., sensors, scenes). Traditional methods, such as costly full retraining or fine-tuning, are inefficient; the latter commonly suffers from catastrophic forgetting on historical data. To overcome this, this research proposes a novel CycleGAN-based incremental SAR image domain adaptation method for Domain Incremental Learning (DIL). First, the SAR Incremental Ship Detection Dataset (SAR-ISDD) was compiled, comprising 31,275 images across six phases, simulating domain shifts via dualdimensional partitioning (sensor type and scene context). The core method uses a CycleGAN to generate transition-domain pseudoimages that fuse features from initial and incremental domains. Integrating this generation with knowledge distillation creates a lightweight DIL framework that effectively mitigates domain shift and preserves knowledge. Experiments show the model significantly improves incremental domain performance (mAP₅₀ of 77.5%, a 4.1 percentage point gain) while successfully minimizing knowledge decay in the initial domain (mAP₅₀ decreased by only 1.9 percentage points), confirming its efficacy for fast, robust model updates.