Multi-unmanned aerial vehicle (Multi-UAV) low-altitude networks require efficient transmission of perception information under limited bandwidth and dynamic channels. Semantic communication is a promising solution. However, existing designs rarely address two critical challenges: Detector confidence is often miscalibrated, and multiple UAVs may transmit redundant semantics from overlapping views. To address these issues, we propose a confidence-calibrated visual semantic communication framework for multi-UAV networks. Each UAV extracts semantic tokens and local detection hypotheses from onboard images. It then estimates calibrated confidence and localization quality for candidate objects. Based on these cues, we define a reliable semantic value. This value integrates calibrated confidence, localization quality, class importance, and cross-UAV semantic novelty. We further formulate a joint wireless-semantic optimization problem covering subchannel assignment, power allocation, semantic token selection, and semantic quantization. To evaluate resource utilization, we introduce confidence-calibrated semantic efficiency (CCSE), which measures reliable semantic gain per unit communication cost. We solve the mixed discrete-continuous stochastic problem by using a hierarchical reinforcement learning-based resource allocation scheme. In this scheme, the upper layer allocates wireless resources, and the lower layer determines semantic content and bit depth. Extended simulation results show that the proposed framework enables efficient and trustworthy multi-UAV visual semantic communications in low-altitude networks and outperforms several baselines.
Zheng et al. (Sat,) studied this question.