Decoding Space Radio Waves: Self‐Supervised AI Deciphers Plasma Frequency

Abstract This article presents a methodology for the extraction of plasma frequency from the upper hybrid resonance traces obtained from space‐based wave spectrograms utilizing self‐supervised artificial intelligence. It elaborates on the development and implementation of a deep learning model. The model utilizes contrastive learning techniques, leveraging positive pairs without incorporating negative pairs. The model's architecture is described, encompassing data augmentation strategies and the employment of a convolutional residual network (ResNet) as the backbone architecture. Two self‐supervised feature representation learning methods, Bootstrap Your Own Latent (BYOL) and PIXel‐level Consistency Learning (PIXCL), are evaluated. The paper additionally examines the integration of a Fully Convolutional Network (FCN) structure for subsequent downstream tasks and the use of semi‐self‐supervised learning to enhance performance with limited labeled data. Results from experiments conducted using data from the Van Allen Probes (VAP) mission demonstrate the efficacy of the proposed method. The model's performance, assessed using the Hausdorff Distance (HD) metric, exhibits promising outcomes in comparison to supervised learning benchmarks, while significantly reducing the necessity for manual data labeling.