Performance comparison of test set.

The classification of hyperspectral images (HSI) is an important foundation in the field of remote sensing. Mamba architectures based on state space model (SSM) have shown great potential in the field of HSI processing due to their powerful long-range sequence modeling capabilities and the efficiency advantages of linear computing. Based on this theoretical basis, We propose a novel deep learning framework: long-sequence Mamba (EchoMamba), which combines the powerful long sequence processing capabilities of Long Short-Term Memory(LSTM) and Mamba to further explore the spectral dimension of HSI, and carry out more in-depth mining and learning of the spectral dimension of HSI. Compared with the previous HSI classification model, the experimental results show that EchoMamba can significantly reduce the training time cost of HSI and effectively improve the performance of the classification task.This study not only advances the current state of HSI classification but also provides a robust foundation for future research in spectral-spatial feature extraction and large-scale remote sensing applications.

高光谱图像（hyperspectral images, HSI）分类是遥感领域的重要基础。基于状态空间模型（state space model, SSM）的Mamba架构，凭借其强大的长程序列建模能力与线性计算的效率优势，在高光谱图像处理领域展现出巨大潜力。基于上述理论基础，本文提出一种新颖的深度学习框架：长序列Mamba（EchoMamba），该框架融合了长短期记忆网络（Long Short-Term Memory, LSTM）与Mamba的优秀长序列处理能力，以进一步探究高光谱图像的光谱维度，并对其光谱维度开展更为深入的挖掘与学习。相较于此前的高光谱图像分类模型，实验结果表明，EchoMamba可显著降低高光谱图像分类的训练时间成本，并有效提升分类任务的性能。本研究不仅推进了当前高光谱图像分类的研究现状，还为未来的光谱-空间特征提取与大规模遥感应用研究提供了坚实的基础。