Autoencoder-Based Deep Learning Approach for 3D Spectrum Completion and Prediction

With the advancement of 6G and the adoption of integrated sensing and communication (ISAC) in low-altitude networks, it is crucial to guarantee the communication performance of ISAC systems by sensing the electromagnetic spectrum situation accurately and promptly. To address this challenge, we propose a 3D spectrum prediction framework for dynamic 3D spectrum allocation. Firstly, we propose an autoencoder-based spectrum completion scheme that recovers missing spectrum data by extracting 3D spatial features from the available data. Our scheme outperforms the benchmark schemes in terms of completion accuracy. Secondly, based on the spectrum completion results, we further propose a deep learning approach for spectrum prediction, which designs specialized branches tailored to different spectrum features to mine the intrinsic correlations. With the better completion results and well-designed prediction network, the prediction performance is significantly improved.