SWiSS: Self-supervised Vision Transformer for Wideband Spectrum Segmentation

Self-supervised learning (SSL) has emerged as an effective approach for capturing robust and generalizable patterns from unlabeled data, which is particularly beneficial in the field of wireless spectrum analysis, where labeled data are often scarce. In this paper, we propose a Transformer-based wideband spectrum segmentation model, termed SWiSS, that leverages masking-based SSL pre-training to learn robust and discriminative spectro-temporal features from unlabeled data. The model is fine-tuned on two downstream tasks, including wireless technology classification (WTC) and signal modulation classification (SMC) and compared with its supervised counterpart under identical settings. To demonstrate the effectiveness of the proposed SWiSS model, we conduct comprehensive experiments on two public datasets, namely SPREAD-small and TorchSig. Experimental results show that SSL pretraining substantially improves convergence stability, validation performance, and segmentation quality compared to the purely supervised model.