LLM-Enhanced Dynamic Spectrum Management for Integrated Non-Terrestrial and Terrestrial Networks: A Multi-Objective Optimization Approach

Efficient spectrum coordination between non-terrestrial and terrestrial networks (NTNs and TNs) is essential to satisfy the throughput and latency targets of next-generation wireless systems. This paper presents LEDSM, a large-language-model-enhanced dynamic spectrum management framework that integrates numerical optimization with linguistic reasoning to handle the heterogeneous, rapidly varying NTN–TN environment. A multi-objective formulation jointly maximizes throughput, minimizes interference, preserves fairness, and improves energy efficiency while maintaining stability under mobility and atmospheric variation. Quantitatively, LEDSM achieves up to 7.2% higher aggregate throughput (61.2 vs. 57.1 Mbps) and 22.4% lower mean interference power (11.8 vs. 15.2 dBm) compared with the best DRL-based baseline in an urban Ka-band scenario. The framework also converges roughly three times faster (5.3 vs. 15.1 iterations to equilibrium) and yields a 2.8% improvement in Jain’s fairness index, without increasing energy consumption. These gains remain consistent across 3GPP-aligned Ku/Ka bands and diverse propagation conditions. The proposed approach therefore provides an interpretable, robust, and computationally efficient path toward intelligent spectrum management in integrated terrestrial–non-terrestrial networks.