<p>How to remove various noises without destroying the edge detail structure is the focus of hyperspectral image (HSI) denoising task. However, existing methods either denoise excessively or insufficiently, resulting in information loss and image distortion. To solve this problem, this paper proposes a wavelet transform guided dual-domain network termed WTGDNet, which applies a wavelet domain and a spatial domain to fully remove the noise of HSIs. In the wavelet domain, we decompose the original noisy HSI into two high-frequency and one low-frequency features by using the wavelet transform to better remove noise in different frequencies. For the high-frequency feature, we first apply the spatial-spectral attention (SSA) to capture the spatial and spectral features, and then apply the feature fusion (FF) to fully exploit the texture and detail information hidden in the noise. In the spatial domain, we directly remove the noise in the original HSI by using the guidance of high-frequency information from the wavelet domain and designing a spatial-spectral quasi-recurrent attention module (SSQRAM) to better capture the spatial information. Notably, SSQRAM is composed of convolution layer, quasi-recurrent pooling, and global channel attention. The quasi-recurrent pooling is used to mine the global correlation along all bands. The global channel attention is used to compute the correlation among different channels, so that the network focuses on noisy channels. Finally, we deploy spatial-spectral residual block (SSRB), which further explores the spatial-spectral correlation while fusing the dual-domain to obtain better denoising output. We perform qualitative and quantitative experiments on publicly available datasets. The results show that WTGDNet outperforms state-of-the-art methods in terms of visualization and objective evaluation metrics.</p>