CA-HLD: context-adaptive cropping with hidden-layer distillation for SAR pretraining
Abstract
Self-supervised learning (SSL) has become a practical paradigm for learning transferable synthetic aperture radar (SAR) representations from abundant unlabeled acquisitions. However, SAR pretraining based on self-distillation with no labels (DINO) leaves two aspects of the training signal insufficiently addressed: random multi-crop may miss or repeatedly sample sparse scattering evidence, and output-level distillation provides no direct supervision for intermediate patch tokens that encode local scattering geometry. To address these limitations, this paper presents CA-HLD, a SAR pretraining framework that couples context-adaptive cropping with hidden-layer distillation. The context-adaptive cropping mechanism constructs a structural context field from multi-scale local mean-ratio responses and selects global and local views around informative, nonredundant scattering regions. In parallel, hidden-layer distillation aligns intermediate teacher and student patch tokens, adding hierarchical supervision to standard self-distillation objectives. CA-HLD improves ATRNet frozen k-NN accuracy by 22.4 percentage points over the strongest baseline and raises SARDet-100K AP from 52.0 to 55.1 under the same detection protocol. These results indicate that view construction and intermediate-layer supervision jointly strengthen SAR representation learning.