IS3 A: An Integrated Soft-Sensing and Spectrum Allocation Scheme in UAV Spectrum Sharing Network

Abstract

Dynamic spectrum access enables unmanned aerial vehicles (UAVs) to opportunistically utilize licensed spectrum. However, conventional architectures decouple spectrum sensing from access decisions, losing soft sensing confidence and introducing cross-layer latency that increase the risk of harmful interference to primary users (PUs). To address these limitations, an integrated soft-sensing and spectrum allocation (IS ${ }^{3}$ A) scheme is proposed. It couples a lightweight micro time-frequency dualstream network (Micro-TFDNet) sensing module with a double deep Q-network with prioritized experience replay (DDQN-PER) decision module. Micro-TFDNet employs a differentiable FFT and adaptive gated fusion to generate soft channel occupancy probabilities directly from raw I/Q samples. These soft probabilities serve as a cross-layer interface that feeds into both the state and the reward, preserving sensing uncertainty and avoiding the information loss caused by hard decisions. Noisy networks and Huber loss enhance exploration and training stability, while prioritized replay emphasizes safety-critical transitions. Simulation results show that Micro-TFDNet achieves the smallest model size and lowest inference latency among baselines, while the IS ${ }^{3}$ A scheme delivers faster convergence and consistently lower primary user collision rates across varying UAV quantities and activity levels.