Multi-Scale Classification of Green Bank Telescope Signals

In this work, we propose a novel pipeline to enhance the detection and mitigation of wide-band Radio Frequency Interference (RFI) in mid-resolution (MR) spectrogram data from the Green Bank Telescope. Our approach integrates an unsupervised Mixture of Experts framework, combining the strengths of multiple edge detection algorithms—Sobel filters, structured forests, Canny edge detection, and the Hough transform—to robustly identify candidate signals in MR data. Leveraging these unsupervised labels, we fine-tune a YOLO-based supervised detection model, significantly enhancing detection efficiency and scalability. Additionally, we embed high-resolution (HR) signals detected by turboSETI into a latent representation space using the Vision Transformer (ViT-B16) model, enabling sophisticated matching between MR and HR signals. This approach substantially reduces false-positive technosignature candidates, improving the efficiency of extraterrestrial signal searches. Our method presents a significant advancement in automated signal detection, laying the groundwork for future large-scale technosignature exploration.