**SSUMamba: Spatial-Spectral Selective State Space Model for Hyperspectral Image Denoising** This paper introduces SSUMamba, a novel approach for hyperspectral image (HSI) denoising that leverages the spatial-spectral continuous scan (SSCS) Mamba block. The method aims to model long-range spatial-spectral correlations in HSI denoising, addressing the challenges posed by noise arising from intra-imaging mechanisms and environmental factors. The SSCS Mamba block alternates row, column, and band in six different orders to generate sequences and uses a bidirectional state space model (SSM) to exploit long-range spatial-spectral dependencies. Additionally, residual 3D convolutions are incorporated to enhance local spatial-spectral modeling. Experimental results on synthetic and real-world datasets demonstrate that SSUMamba achieves superior denoising performance with lower memory consumption per batch compared to transformer-based methods. The source code is available at https://github.com/Ironkitty/SSUMamba. **Key Contributions:** - Introduction of the spatial-spectral U-Mamba (SSUMamba) for HSI denoising. - Development of the spatial-spectral continuous scan (SSCS) Mamba block to generate continuous sequences and model long-range spatial-spectral dependencies. - Incorporation of residual 3D convolutional blocks to enhance local spatial-spectral modeling. - Superior denoising performance with fewer parameters and lower memory consumption per batch compared to transformer-based methods. **Method Overview:** - SSUMamba is designed to learn the mapping from noisy HSI to clean HSI. - The model consists of a feature extractor, encoder blocks, SSCS Mamba blocks, decoder blocks with skip connections, and a reconstructor. - The SSCS Mamba block uses a bidirectional SSM with SSCS to capture both local and global spatial-spectral correlations and continuity. - The model is trained using a loss function that minimizes the $L_2$ distance between the predicted clean HSI and the ground truth clean HSI. **Experimental Results:** - SSUMamba outperforms other methods in terms of PSNR, SSIM, and SAM metrics on synthetic datasets. - It demonstrates superior denoising performance on real-world HSIs, including the Gaofen-5 Wuhan HSI and Earth Observing-1 HSI. - Ablation studies show the effectiveness of the SSCS scan schemes, bidirectional SSM, and residual blocks. - Network width experiments confirm that increasing the network width improves performance, with SSUMamba achieving better results compared to transformer-based methods in terms of parameters, memory consumption, and inference time. **Conclusion:** SSUMamba effectively models long-range spatial-spectral correlations in HSI denoising, achieving state-of-the-art performance with reduced computational complexity. Future work will explore more efficient network designs and integrate the SSCS Mamba block into other tasks.**SSUMamba: Spatial-Spectral Selective State Space Model for Hyperspectral Image Denoising** This paper introduces SSUMamba, a novel approach for hyperspectral image (HSI) denoising that leverages the spatial-spectral continuous scan (SSCS) Mamba block. The method aims to model long-range spatial-spectral correlations in HSI denoising, addressing the challenges posed by noise arising from intra-imaging mechanisms and environmental factors. The SSCS Mamba block alternates row, column, and band in six different orders to generate sequences and uses a bidirectional state space model (SSM) to exploit long-range spatial-spectral dependencies. Additionally, residual 3D convolutions are incorporated to enhance local spatial-spectral modeling. Experimental results on synthetic and real-world datasets demonstrate that SSUMamba achieves superior denoising performance with lower memory consumption per batch compared to transformer-based methods. The source code is available at https://github.com/Ironkitty/SSUMamba. **Key Contributions:** - Introduction of the spatial-spectral U-Mamba (SSUMamba) for HSI denoising. - Development of the spatial-spectral continuous scan (SSCS) Mamba block to generate continuous sequences and model long-range spatial-spectral dependencies. - Incorporation of residual 3D convolutional blocks to enhance local spatial-spectral modeling. - Superior denoising performance with fewer parameters and lower memory consumption per batch compared to transformer-based methods. **Method Overview:** - SSUMamba is designed to learn the mapping from noisy HSI to clean HSI. - The model consists of a feature extractor, encoder blocks, SSCS Mamba blocks, decoder blocks with skip connections, and a reconstructor. - The SSCS Mamba block uses a bidirectional SSM with SSCS to capture both local and global spatial-spectral correlations and continuity. - The model is trained using a loss function that minimizes the $L_2$ distance between the predicted clean HSI and the ground truth clean HSI. **Experimental Results:** - SSUMamba outperforms other methods in terms of PSNR, SSIM, and SAM metrics on synthetic datasets. - It demonstrates superior denoising performance on real-world HSIs, including the Gaofen-5 Wuhan HSI and Earth Observing-1 HSI. - Ablation studies show the effectiveness of the SSCS scan schemes, bidirectional SSM, and residual blocks. - Network width experiments confirm that increasing the network width improves performance, with SSUMamba achieving better results compared to transformer-based methods in terms of parameters, memory consumption, and inference time. **Conclusion:** SSUMamba effectively models long-range spatial-spectral correlations in HSI denoising, achieving state-of-the-art performance with reduced computational complexity. Future work will explore more efficient network designs and integrate the SSCS Mamba block into other tasks.