【摘要】 阐述一种基于条件扩散模型的智能电磁频谱图重构方法，通过将区域海拔、环境表面参数、发射机天线参数和位置等信息作为生成模型条件，实现经济高效的频谱图生成，尤其适用于复杂场景。对于难以直接嵌入图像通道的语义特征，如目标类型和重建方式，引入大规模视觉语言模型CLIP进行跨模态预训练，并通过交叉注意力将先验特征整合到扩散模型中。同时，将文本提示作为生成条件，构建双重条件扩散框架以增强下游任务处理能力。通过使用射线追踪方法合成收集训练数据，覆盖100～10 000MHz频段、10～50dB增益及各类郊区开阔地形，包括全向通信与定向波束雷达目标。全面评估表明，该方法能够从地理环境和相关参数重建电磁频谱图，且预测准确率达到95%，超过现有SOTA方法。更多还原

【Abstract】 This paper describes an intelligent electromagnetic spectrum reconstruction method based on conditional diffusion model, which achieves cost-effective spectrum generation by using information such as regional altitude, environmental surface parameters, transmitter antenna parameters, and location as generation model conditions, especially suitable for complex scenarios. For semantic features that are difficult to directly embed into image channels, such as target types and reconstruction methods, it introduces a large-scale visual language model CLIP for cross modal pre training, and integrates prior features into the diffusion model through cross attention. At the same time, it uses text prompts as generation conditions and constructs a dual conditional diffusion framework to enhance downstream task processing capabilities. By using ray tracing methods to synthesize and collect training data, covering the frequency range of 100-1000 MHz, 10-50 dB gain, and various suburban open terrain, including omnidirectional communication and directional beam radar targets. A comprehensive evaluation shows that this method can reconstruct electromagnetic spectrum maps from geographical environments and related parameters, with a prediction accuracy of 95%, surpassing existing SOTA methods.更多还原