Abstract: An enhanced image recognition algorithm for Auroral Kilometric Radiation (AKR) detection is presented by integrating a self-attention mechanism into the VGG16 Convolutional Neural Network (CNN) architecture. The primary goal is to improve the flexibility and detection accuracy of AKR identification, which is crucial for understanding the dynamic changes in Earth’s radiation belts and the associated energetic particle variations. The methodology begins with employing the VGG16 CNN as the foundational model to extract local features from raw data that are instrumental in AKR recognition. Subsequently, a custom Self-Attention Mechanism (SAM-V) is embedded in the VGG network. The Self-Attention Mechanism (SAM), originally designed for sequential data processing, is adapted to work with the VGG16 network. Traditional integration of SAM with VGG16 could potentially increase the model’s complexity and computational cost, leading to potential feature sparsity issues. However, the proposed custom SAM-V generates queries, keys, and values through defined convolutional layers, offering more control over feature input and output. This customization implies shared parameters, reducing the number of model parameters, thereby mitigating the risk of overfitting and enhancing the model’s generalization capabilities. This approach is particularly adept at capturing correlations in power spectral density across different time points or frequencies, minimizing the impact of noise and improving recognition accuracy. Additionally, a linear learning rate warm-up and dynamic decay strategy are employed to accelerate model convergence and enhance generalization. The experimental results demonstrate that the improved model achieves an average recognition accuracy of approximately 93%, which represents a 3.3% increase compared to the original VGG16 model. Furthermore, other performance metrics such as recall rate and precision have also seen significant improvements. In conclusion, the integration of a custom self-attention mechanism into the VGG16 network has yielded a more efficient and accurate model for AKR detection. This advancement not only bolsters the study of auroral kilometric radiation but also has broader implications for the analysis of Earth’s radiation belt dynamics and energetic particle behavior. The model’s enhanced generalization capabilities and improved accuracy underscore the potential for applying similar techniques to other image recognition tasks within the field of space physics and beyond.