珊瑚礁的科学管理具有重要的生态与经济意义。水下高光谱影像（UHI）可实现精细尺度的海底生境分类与制图，为珊瑚礁监测与保护提供有效支撑。然而，UHI 的高维特性、珊瑚礁海底生境类型的显著空间异质性，以及有限的训练样本，共同制约了判别性空-谱特征的有效提取，进而影响细粒度海底生境分类性能。此外，不同海底生境类型在空间分布与信息密度上的不均衡性，导致特征学习偏向显著区域，削弱模型对小尺度或弱响应类别的判别能力。为应对上述局限，本文提出一种基于 Mamba 架构的珊瑚礁水下高光谱图像分类模型——CRUHI-Mamba。首先，构建空-谱 Mamba 主干网络（SSMB），通过双池化层与空-谱双分支 Mamba 模块的协同设计，实现对全局空间依赖关系与细粒度光谱差异的联合建模与特征提取，即使在训练样本有限条件下亦能保持有效性。其次，为缓解因海底生境类型间信息密度不均所引发的分类偏差，设计自适应特征-决策融合模块（AFDFM），动态平衡空-谱特征信息，防止信息丢失，增强特征表征能力，并校正分类偏差，最终生成鲁棒的分类图。在三个不同水深采集的珊瑚礁 UHI 数据集上开展的分类实验表明，CRUHI-Mamba 在定量精度方面优于当前主流方法。

Scientific management of coral reefs is of significant ecological and economic importance. Underwater hyperspectral imagery (UHI) enables fine-scale benthic classification and mapping, providing effective support for coral reef monitoring and conservation. However, the high-dimensional nature of UHI and the pronounced spatial heterogeneity of coral reef benthic types, coupled with limited training samples, hinder the effective extraction of discriminative spatial–spectral features for fine-grained benthic classification. Moreover, the imbalance in spatial distribution and information density across benthic types biases feature learning toward prominent regions, thereby compromising the model’s discrimination of small-scale or weak-response types. To address these limitations, a coral reef underwater hyperspectral image classification model based on the Mamba architecture, termed CRUHI-Mamba, is proposed. First, a spatial–spectral Mamba backbone (SSMB) is constructed, which, through the synergistic design of dual pooling layers and spatial–spectral dual-branch Mamba blocks, enables the collaborative modeling and feature extraction of global spatial dependencies and fine-grained spectral differences, even with limited training samples. Next, to mitigate classification bias caused by the uneven information density among benthic types, an adaptive feature–decision fusion module (AFDFM) is designed to dynamically balance spatial–spectral feature information, prevent information loss, enhance feature representations, and correct classification bias, ultimately generating robust classification maps. Classification experiments on coral reef UHI data collected at three different water depths demonstrate that CRUHI-Mamba outperforms state-of-the-art methods in quantitative accuracy, computational complexity, and efficiency, achieving an overall accuracy exceeding 99.22% with an inference time of less than 10 seconds. CRUHI-Mamba therefore provides an effective methodological support for fine-scale benthic habitat mapping in coral reefs.