Manganese nodules, widely distributed across the deep-sea floor, are emerging as a significant potential mineral resource. Addressing the need for improved exploration and classification methods, this study explores the application of underwater hyperspectral imaging technology for the detection and classification of manganese nodule ore and other rock types. The system, operating within a spectral range of 400 - 1000 nm and achieving a spectral resolution of less than 5 nm, captures the spectral characteristics and spatial information of manganese nodule ores. Different classifiers, including Spectral Angle Mapper (SAM), Support Vector Machine (SVM), and Convolutional Neural Networks (2D-CNN and 3DCNN) to analyze the spectral data. Our results indicate that the four classifiers can effectively achieve ore classification, and CNN-based classifiers significantly outperform traditional SAM and SVM methods. The 2D-CNN model achieved the highest OA at 92.27%, closely followed by the 3D-CNN model at 91.77%. Our findings demonstrate the potential of underwater hyperspectral imaging combined with advanced machine learning techniques in marine mineral detection and environmental monitoring.