Accurate automatic modulation recognition (AMR) of digital M-ary signals remains a fundamental yet challenging task in wireless communication networks, particularly under low signal-to-noise ratio (SNR) conditions. Conventional approaches, including maximum likelihood estimation, decision-theoretic methods, and classical pattern recognition techniques, often lead to limited robustness and adaptability in dynamic propagation environments. To address these limitations, an enhanced modulation learning and classification algorithm (EMLCA) was introduced, which utilizes a hybrid architecture that integrates convolutional neural networks and long short-term memory networks. A Mask_1/residual network (ResNet)-based feature enhancement strategy was incorporated to improve resilience, while Aquila optimization was employed to adaptively tune network parameters and enhance classification stability. Model training was guided by a reduced-loss formulation combining cross-entropy and mean squared error (MSE) objectives. Comprehensive simulations were conducted across multiple feature domains, including statistical, wavelet, and spectral representations, under SNR conditions ranging from $-$20 dB to 20 dB. The obtained results demonstrate that EMLCA consistently outperformed conventional AMR methods in terms of recognition accuracy, computational efficiency, and adaptability. A maximum recognition accuracy of approximately 95% was achieved under challenging noise conditions, accompanied by a reduction in processing time of nearly 25% relative to benchmark techniques. Furthermore, adaptability analysis confirms that the proposed framework maintained stable performance under varying channel distortions and environmental dynamics. These findings indicate that EMLCA provides a robust and scalable solution for real-time modulation recognition and offers strong potential for deployment in adaptive and next-generation wireless communication systems.