矿物光谱综合反映了岩矿的物理化学特性、 组分和内部结构特征, 已被应用于岩矿识别研究。 传统的矿物光谱分类方法需要先对矿物光谱进行预处理, 再采用不同方法分析光谱特征, 从而实现分类目的。 但同时也会造成部分光谱信息丢失, 导致最终分类精度不高且操作过程繁琐、 效率低下, 难以应对日益增长的大数据处理需求。 因此, 建立一个准确、 高效的矿物光谱自动分类模型意义重大。 卷积神经网络是应用最广泛的深度学习模型之一, 它通过逐层抽取数据特征并组合形成高层语义信息, 具有极强的模型表达能力, 在光谱数据分析方面应用潜力巨大。 针对矿物光谱数据的特点, 提出了基于一维空洞卷积神经网络(1D-DCNN)的矿物光谱分类方法, 利用空洞卷积神经网络提取光谱特征, 采用反向传播算法结合随机梯度下降优化器调整模型参数, 输出光谱分类结果, 实现了矿物类别的端到端检测。 该网络包含1个输入层、 3个空洞卷积层、 2个池化层、 2个全连接层和1个输出层, 采用交叉熵为损失函数, 引入空洞卷积扩大滤波器感受野, 有效避免光谱细节特征丢失。 实验采集了白云母、 白云石、 方解石、 高岭石四种矿物光谱, 并通过添加噪声的方式进行数据增强, 构建数量充足的矿物光谱样本用于神经网络模型训练与测试; 探讨了卷积类型、 迭代次数对模型分类结果的影响, 并与多种传统矿物光谱分类方法进行对比, 评价模型性能。 实验结果表明, 提出的1D-DCNN模型可实现矿物光谱快速准确分类, 分类准确率达到99.32%, 优于反向传播算法(BP)和支持向量机(SVM), 说明所提方法能够充分学习矿物光谱特征并有效分类, 且模型具有良好的鲁棒性和可扩展性。 该方法也可推广到煤炭、 油气、 月壤等其他领域光谱分类应用中。

The spectrum is a comprehensive reflection of the mineral’s physical chemistry characteristics, composition and structure, which has been used in mineral and rock identification. The traditional classification methods of the mineral spectrum require complex spectral pretreatment, and then some spectral features are analyzed by different methods to achieve the goal of fine classification. However, the pretreatment may cause a partial loss of the spectral information and reduce the classification accuracy. Besides, the operation process is complex, so the efficiency is low, making it difficult to cope with the growing demand for big data processing. Therefore, it is important to establish an accurate, efficient and automatic classification model for the mineral spectrum. As one of the widely used deep learning models, the convolutional neural network extracts data features layer by layer and combines them to form higher-level semantic information. It has a strong capability of model formulation and great potential for the analysis of spectral data. This paper proposes a novel mineral spectrum classification method based on a one-dimensional dilated convolutional neural network (1D-DCNN). The DCNN is used for spectral feature extraction. The backpropagation algorithm combined with the random gradient descent optimizer is used to adjust the model’s parameters, then output the classification result, which implements the end-to-end discrimination of mineral species. The 1D-DCNN includes one input layer, three dilated convolution layers, two pooling layers, two full connection layers and one output layer. It uses cross-entropy as the loss function, and dilated convolution is introduced to enlarge the receptive field of filters effectively avoid the loss of spectral feature details. The spectrum of four different minerals, muscovite, dolomite, calcite and kaolinite, are collected, and the data are augmented by way of adding noise to construct sufficient spectral samples, which are used for model training and testing. Then, we explore the impacts of different model parameters, such as the convolution type and the number of iterations, and then compare the proposed model with the traditional mineral spectrum classification methods to evaluate its performance. Experimental results indicate that the 1D-DCNN model can quickly and accurately classify mineral spectrum with the accuracy of 99.32%, which is superior to the backpropagation (BP) algorithm and support vector machine (SVM) methods, and it shows that the proposed method can fully learn mineral spectral features and implement a fine classification result, with good robustness and scalability. The proposed method can apply further to the spectra classification in coal, oil-gas, lunar soil and other fields.