DINOSAR: large-scale Sar self-supervised pretraining with content-aware view construction

Self-supervised learning (SSL) provides a scalable approach to learning transferable representations from unlabeled synthetic aperture radar (SAR) imagery. A common SSL strategy is label-free self-distillation, which derives supervision by aligning randomly cropped global and local views. However, in SAR imagery, informative scattering structures are sparsely and unevenly distributed, so random view construction may miss target content or repeatedly sample the same strong scatterers, weakening the consistency signal. To address these challenges, we propose DINOSAR, a SAR self-supervised pretraining framework with content-aware multi-crop (CAMC). CAMC derives a structural content map from multi-scale ratio-gradient responses, extracts scene-adaptive anchors, and selects global and local views by balancing scattering strength, anchor proximity, and coverage diversity. We also construct UniSAR-7M, a multi-source corpus of approximately seven million SAR pretraining samples assembled from public datasets and openly released commercial satellite imagery. Across classification, object detection, semantic segmentation, and cross-modal retrieval, DINOSAR achieves strong transfer performance, including 77.02% frozen k-NN accuracy on ATRNet-STAR SOC-40 and 55.28 AP on SARDet-100K. Controlled comparisons against a matched Random-crop DINO baseline and a SAR-1M-only variant further confirm the benefits of content-aware view construction and of UniSAR-7M’s scale and diversity, respectively.