To mitigate the adverse effects of strong interference on the data quality and detection performance of electromagnetic methods, and to address the suboptimal signal-noise separation performance resulting from inappropriate parameter settings in deep learning approaches. This paper proposes an electromagnetic signal-noise identification method based on deep learning optimized by the Hybrid Strategy Improved Hippopotamus Optimization (HSIHO) algorithm. First, multiple optimization strategies are evaluated through comparative analysis, and a hybrid strategy combining lens opposite-based learning and adaptive T-distribution perturbation is selected to enhance the original hippopotamus optimization algorithm. Subsequently, the improved HSIHO algorithm is employed to optimize key parameters of the deep learning model, including the learning rate, number of neurons, and number of iterations. Finally, the optimized model is applied to electromagnetic signal-noise identification, and its performance is compared with that of other deep learning models. Experimental results demonstrate that, in benchmark function tests, the proposed HSIHO algorithm exhibits superior convergence speed, solution accuracy, flexibility, and scalability compared to other intelligent optimization algorithms. In practical applications involving electromagnetic data processing, the HSIHO-optimized deep learning model achieves significantly higher accuracy in signal-noise identification than probabilistic neural networks, convolutional neural networks, long short-term memory networks, and gated recurrent unit. The processed electromagnetic data show enhanced quality and greater stability in electric field curve morphology. The proposed method provides effective technical support for the inversion and interpretation of electromagnetic survey data.