SpectralMamba is a novel efficient deep learning framework for hyperspectral (HS) image classification, designed to address the inefficiencies of traditional recurrent neural networks (RNNs) and Transformers. These models, while effective in capturing long-range dependencies in spectral sequences, suffer from computational inefficiencies due to their reliance on attention mechanisms and sequential processing. SpectralMamba integrates a state space model (SSM) with an efficient deep learning framework, enabling it to handle HS data more effectively. The framework operates in two levels: first, in the spatial-spectral space, it uses efficient convolutions to learn a dynamic mask that encodes spatial regularity and spectral peculiarity, reducing spectral variability and confusion. Second, the merged spectrum is processed in the hidden state space with parameters learned input-dependently, allowing for selective responses without the need for attention or recurrence. A piece-wise scanning mechanism is employed to convert continuous spectra into sequences with reduced length, maintaining contextual profiles. Extensive experiments on four benchmark HS datasets (Houston2013, Augsburg, Longkou, Botswana) show that SpectralMamba outperforms existing methods in both performance and efficiency. It achieves high accuracy with significantly lower computational costs, demonstrating its effectiveness in HS image classification. The framework's key components include piece-wise sequential scanning (PSS) and gated spatial-spectral merging (GSSM), which enhance the model's ability to handle spectral variability and improve discriminative representation. SpectralMamba's architecture is lightweight and efficient, with a simplified model that reduces computational overhead. It is capable of handling both pixel-wise and patch-wise inputs, making it versatile for various applications. The model's performance is validated through extensive experiments, showing its superiority in classification accuracy and efficiency compared to traditional methods. The framework's ability to adapt to different HS data characteristics and its efficient computation make it a promising solution for HS image classification in remote sensing applications.SpectralMamba is a novel efficient deep learning framework for hyperspectral (HS) image classification, designed to address the inefficiencies of traditional recurrent neural networks (RNNs) and Transformers. These models, while effective in capturing long-range dependencies in spectral sequences, suffer from computational inefficiencies due to their reliance on attention mechanisms and sequential processing. SpectralMamba integrates a state space model (SSM) with an efficient deep learning framework, enabling it to handle HS data more effectively. The framework operates in two levels: first, in the spatial-spectral space, it uses efficient convolutions to learn a dynamic mask that encodes spatial regularity and spectral peculiarity, reducing spectral variability and confusion. Second, the merged spectrum is processed in the hidden state space with parameters learned input-dependently, allowing for selective responses without the need for attention or recurrence. A piece-wise scanning mechanism is employed to convert continuous spectra into sequences with reduced length, maintaining contextual profiles. Extensive experiments on four benchmark HS datasets (Houston2013, Augsburg, Longkou, Botswana) show that SpectralMamba outperforms existing methods in both performance and efficiency. It achieves high accuracy with significantly lower computational costs, demonstrating its effectiveness in HS image classification. The framework's key components include piece-wise sequential scanning (PSS) and gated spatial-spectral merging (GSSM), which enhance the model's ability to handle spectral variability and improve discriminative representation. SpectralMamba's architecture is lightweight and efficient, with a simplified model that reduces computational overhead. It is capable of handling both pixel-wise and patch-wise inputs, making it versatile for various applications. The model's performance is validated through extensive experiments, showing its superiority in classification accuracy and efficiency compared to traditional methods. The framework's ability to adapt to different HS data characteristics and its efficient computation make it a promising solution for HS image classification in remote sensing applications.