Normality-Calibrated Modality Routing for Optical-SAR Object Detection Under Cloud-Induced Optical Degradation

Optical-SAR object detectors trained on cloud-free paired imagery tend to develop a strong dependence on the information-rich optical modality. Consequently, the availability of SAR observations does not necessarily guarantee their effective use when optical information becomes unreliable under cloud occlusion. To address this issue, we propose NCMR-Net, a Normality-Calibrated Modality Routing Network that enables spatially adaptive optical-to-SAR routing without image-level cloud augmentation. NCMR-Net retains the standard YOLOv11s detection architecture and introduces three functional modules along the optical-SAR interaction and routing pathway. First, the Frequency-Selective SAR Refinement (FSR) module applies Haar-based subband decomposition and bounded high-frequency correction to construct a more usable SAR fallback representation. Second, the Normality-Calibrated Gating (NCG) module derives non-learnable input-domain evidence from high brightness and reduced local texture, which characterize the optical information collapse associated with cloud degradation. NCG calibrates this evidence against its empirical distribution under normal optical observations. At inference, a lower bound on the SAR mixing coefficient maintains optical-SAR fusion at low NCG responses, while stronger responses increase the SAR coefficient above this bound. Finally, the Gate-Conditioned Residual Adaptation (GCRA) module combines the routing gate with local feature statistics and performs gate-conditioned, identity-initialized residual adaptation across different routing states. In the single-run OSPRC evaluation, NCMR-Net is trained on cloud-free RGB10M-VH pairs without image-level cloud augmentation. Under the matched RGB10M-VH condition, NCMR-Net achieves 85.12%mAP50 and 53.57%mAP50:95, compared with 83.48% and 53.30% for YOLOv11s-Add. Without retraining, fine-tuning, or test-time adaptation, it achieves 64.48%mAP50 under the synthetic cloud-degraded CLUDE-VH condition, exceeding the baseline’s 39.15% by 25.33 percentage points. The corresponding mAP50 retention rates are 75.8% for NCMR-Net and 46.9% for YOLOv11s-Add. These results support improved performance retention under the tested synthetic cloud degradation; generalization to real clouds remains to be established.