Hybrid Deep Learning-Based Adaptive Modulation for Free Space Optical (FSO) Communication

Throughput values are closer to the required values of some current wireless systems, and free-space optics (FSO) has several promising applications in scenarios where high bandwidth demand necessitates a low-cost electronic warfare (EW) strategy. The influence of atmospheric turbulence, weather variability, and pointing errors results in severe impairment. Conventional adaptive modulation schemes (most of which are rule-based or based on shallow learning) may not adapt well to variable and nonlinear channel conditions. In this paper, we introduce a new adaptive modulation scheme that uses a mix of deep learning techniques, including CNNs, LSTM units, and an attention mechanism, to classify real time modulation schemes in free-space-optical (FSO) communication networks. The model captures the spatiotemporal features of received optical signals and dynamically depicts modulation formats based on the channel states. Simulation results demonstrate a classification accuracy performance of 96.3%, a Bit Error Rate (BER) reduction rate exceeding 30% compared to the fixed scheme, and a maximum average spectral efficiency gain of 45% even under the worst scenarios. Furthermore, with an adaptation latency of ~48ms, the proposed system enables real-time service in dynamic and mobile environments. The investigation on the attention maps also demonstrates the explainability and multi-reasoning properties of our model. This hybrid architecture is a scalable and reconfigurable paradigm for the future optical wireless communications systems deployed in harsh conditions such as UAV channels, smart environments including cities, and inter-satellite communication links.