Abstract: The proliferation of dynamic network applications, from intelligent transportation systems to industrial IoT, has exposed the critical limitations of traditional static and reactively adaptive network topologies. These legacy systems, characterized by siloed optimization and a lack of cross-layer awareness, struggle to meet the stringent demands for ultra-low latency, high reliability, and efficient resource utilization. This paper proposes a novel framework for an Adaptive Network Topology based on Real-Time Multi-Spectral Analysis to address this fundamental challenge. Our solution introduces a hierarchical architecture integrating a Multi-Spectral Fusion and Prediction (MSFP) algorithm, which uses a hybrid Graph Attention Network and Gated Recurrent Unit (GAT-GRU) model to create a predictive state of the network. It also uses the Predictive Spectral-Topology Co-optimization (PSTC) algorithm, a Deep Reinforcement Learning (DRL) agent that prescriptively co-optimizes spectrum access and network routing. Evaluated through extensive simulations in NS-3 across urban vehicular (VANET), satellite, and industrial optical wireless sensor network (OWSN) scenarios, the proposed framework demonstrates statistically significant performance gains over state-of-the-art baselines. Results show an average improvement of 6.6% in Packet Delivery Ratio (PDR), a 27.5% reduction in end-to-end latency, a 14.7% increase in spectrum utilization efficiency, and a 28.5% reduction in control overhead. This research conclusively demonstrates that the synergistic, predictive co-optimization of the physical and network layers is a paradigm shift essential for building the autonomous, high-performance networks of the future.

Key words: real-time multi-spectral analysis, deep reinforcement learning (DRL), software-defined networking (SDN), predictive analytics, graph neural networks (GNN)