How does your GAN-based approach specifically address the small-sample problem better than conventional CNN-based classifiers?
Thank you for your question. In our GAN-based approach, the small-sample problem is addressed through a data augmentation mechanism inherent to the adversarial training process. Unlike conventional CNN-based classifiers, which rely solely on the original training set and are prone to overfitting when data is limited, our method leverages the generator network to synthesize diverse and realistic power spectral signals that expand the effective training distribution. Specifically, the generator learns to produce terrain-like echo spectra from random noise, guided by the discriminator’s feedback. This process not only increases the quantity of training samples but also enriches their feature diversity, helping the model generalize better to unseen data. Moreover, the dual-probe discriminator structure in our framework simultaneously performs discrimination and classification, enabling the model to learn shared feature representations that are robust to data scarcity. In contrast, conventional CNNs typically require large labeled datasets to learn generalizable features and often struggle with limited samples, even when data augmentation techniques like rotation or noise injection are applied. Our GAN-based framework, by contrast, generates semantically meaningful and domain-relevant samples, which leads to improved classification performance and reduced overfitting, as evidenced by the high accuracy (over 98%) achieved across all terrain categories in our experiments.
