Gamma–hadron discrimination remains a fundamental challenge in very-high-energy gamma-ray astronomy due to the strong overlap between gamma-ray–initiated and hadron-induced air showers recorded by imaging atmospheric Cherenkov telescopes, particularly at low energies where background contamination is severe. Traditional cut-based and non-optimized machine learning approaches often struggle to fully exploit the nonlinear and correlated nature of Cherenkov image parameters, leading to suboptimal background suppression and reduced telescope sensitivity. To address these limitations, this paper proposes a unified deep learning and metaheuristic optimization framework that combines an enhanced attention-based long short-term memory network (EALSTM) with advanced optimization strategies. In particular, a novel Adaptive Balanced Greylag Goose Optimization algorithm (ABGGO) is employed to jointly perform feature selection and hyperparameter optimization, enabling effectiveexploration–exploitation balancing while preserving physically meaningful feature representations. The proposed ABGGO+EALSTM framework is systematically evaluated against baseline deep learning models, including artificial neural networks (ANN), convolutional neural networks (CNN), and standard long short-term memory networks (LSTM), under identical experimental conditions. Experimental results on a Monte Carlo–generated Cherenkov telescope dataset demonstrate clear and consistent performance gains at every stage of the analysis. In the baseline evaluation stage, EALSTM achieves an accuracy of 0.9294 and an F-score of 0.9266, outperforming ANN, CNN, and LSTM. Following metaheuristic optimization, the proposed ABGGO+EALSTM model attains a peak accuracy of 0.9718, sensitivity of 0.9694, specificity o f 0 .9740, a nd F-score o f 0 .9705, representing absolute improvements exceeding 4% over the baseline EALSTM configuration and outperforming GA+EALSTM, GWO+EALSTM, and PSO+EALSTM variants. These results demonstrate that integrating attention-based deep learning with adaptive metaheuristic optimization significantly enhances gamma–hadron discrimination, leading to improved background suppression and signal retention. The proposed framework offers a scalable and robust solution for current and next-generation Cherenkov observatories, with strong potential for real-time event filtering, multi-telescope analysis, and future deployment on real observational data.