Machine Learning based Channel Estimation for Gamma-Gamma Noise

We study pilot-aided channel estimation for free-space optical (FSO) links operating under atmospheric turbulence and intentional interference. Both the legitimate and jamming paths experience Gamma-Gamma fading, while the jammer follows an on-off (Bernoulli) activity model that yields non-Gaussian, impulsive disturbances at the receiver. Building on an exact likelihood for this setting, we derive a steepest-descent maximum-likelihood (ML) estimator for the main channel coefficient and extend it to a maximum a posteriori (MAP) form that incorporates Gamma-Gamma priors. We also explore a lightweight hybrid scheme that refines the steepest-descent algorithm (SDA) output with a shallow neural corrector. Monte-Carlo simulations across multiple turbulence conditions (α,β), signal-to-noise ratios (SNRs), signal-to-jamming ratios (SJRs), and jammer activity probabilities demonstrate that the proposed estimators achieve consistently lower mean square error (MSE) and Bit Error Rate (BER) than conventional mean-based or Gaussian-assumed baselines, with graceful degradation as jamming becomes more frequent and clear gains as SJR improves. The results highlight that respecting the true Gamma-Gamma statistics rather than relying on Gaussian surrogates materially improves estimation robustness in hostile FSO environments, while the gradient-based forms remain simple enough for practical implementation.