Wavelet-enhanced cGAN with frequency-spatial domain collaborative perception for celestial spectral denoising

Abstract Doppler velocimetry navigation, as an emerging high-precision celestial navigation technique, determines the radial velocity of spacecraft by analyzing the Doppler shift in spectral lines. However, despite its great potential, the accuracy of current methods remains greatly limited due to insufficient consideration of celestial spectral noise interference. To tackle this issue, we propose an innovative denoising methodology that employs a Wavelet-enhanced conditional generative adversarial network with frequency-spatial domain collaborative perception. Central to this approach is the dual-tree wavelet transform, which enables multiscale spectral decomposition. This process not only facilitates precise frequency-domain feature extraction but also enhances the retention of fine detail information. Furthermore, a novel conditional multi-head Attention mechanism is incorporated. This mechanism establishes long-range dependencies within spectral data and improves global context awareness through adaptive feature recalibration. Our methodology effectively suppresses noise in spectral information while preserving critical spectral features and fine detail information, thereby ensuring accurate reconstruction of spectral data. Experimental results demonstrate that our method manifests robust denoising performance across varying noise levels and diverse spectral data. More specifically, it successfully reduces spectral angle mapper to a mere 0.01 rad and elevates peak signal-to-noise ratio to 39.1 dB in high-noise scenarios ( σ ⩾ 0.5). These advantages highlight the strong spectral denoising capability of our approach, which enhances the accuracy and stability of Doppler velocimetry.