Accuracy and Computational Cost of Machine Learning Models for Electromagnetic Object Classification

Classification of randomly oriented objects from electromagnetic data is important in applications such as remote sensing and security screening, where robustness to simulation/measurement configuration and data variability is critical. While prior work has shown that attention mechanisms can significantly improve classification performance for long, highly structured electromagnetic scattering sequences, their benefit in simpler data regimes remains unclear. In this numerical work, we study classification using plane-wave excitation and electromagnetic field data recorded by a limited number of receiver antennae, resulting in short input sequences. We show that under these conditions, a simple one-dimensional convolutional neural network achieves nearly the same accuracy as attention-based models. We further compare neural networks with classical machine learning methods, including support vector machines (SVMs) and extreme gradient boosting (XGB), and demonstrate that for large datasets these methods perform comparably to neural networks, whereas neural networks offer a clear advantage in the small-data regime. However, neural networks’ tuning and training times can be substantially higher than those of SVMs and XGB. These results provide practical guidance on selecting model complexity for electromagnetic classification based on dataset size and structure, computational resources, and time.