Scattering-Topology Guided and Evidential Uncertainty-Aware Dual-Stream YOLO for Optical-SAR Remote Sensing Object Detection

Abstract

Optical and synthetic aperture radar (SAR) imagery are naturally complementary for remote sensing object detection, yet existing dual-stream YOLO frameworks estimate fusion weights heuristically, confine the SAR scattering mechanism within the SAR branch, and keep frequency-domain priors at the image fusion stage. This article proposes STEU-YOLO, a dual-stream detector built on three coupled modules. The Scattering-Topology Guided Cross-modal Semantic Bridge extracts attribute scattering centers from the SAR branch, builds a $k$ -nearest-neighbor topology graph, and uses graph attention to generate a physical prior that inversely modulates the optical features. The Evidential Uncertainty-Aware Fusion module places a Dirichlet distribution on each modality at each spatial location, so the fusion weight becomes the normalized reciprocal of the total uncertainty and carries Bayesian semantics. The Dual-domain Frequency-Spatial Collaborative Neck moves wavelet decomposition into the detector neck, separating high-frequency scattering from low-frequency structural cues on multi-scale features. On OGSOD-1.0, STEU-YOLO reaches 86.4% mAP50 with 4.72 M parameters, 13.2 GFLOPs, and 87 FPS, exceeding the strongest baseline by 3.5 points. Zero-shot transfer to OGSOD-2.0 retains 93.5% of the accuracy, and the lead widens on the low-resolution, cross-polarization, and open-sea subsets of M4-SAR.