自然界中岩石一般是由多种矿物集合而成的紧致混合物, 由于高光谱传感器低空间分辨率的特征, 获得的高光谱数据多为矿物组分的综合反映。 受噪声干扰以及矿物复杂的混合机理等因素影响, 高光谱端元识别和定量分析成为目前研究的热点与难点。 基于深度学习理论, 对原始自编码结构进行改进, 提出了一种改进的堆栈稀疏自编码的矿物高光谱端元识别方法（stacked sparse autoencoders, SSAE）, 为高光谱解混提供新的思路。 首先, 根据矿物混合光谱的特点, 对原始自编码结构进行三点改进: 第一, 去掉自编码神经网络的偏置项（bias）; 第二, 在隐藏层激活函数之前添加批归一化（batch normalization, BN）层, 最后一层输出层使用Relu激活函数; 第三, 用光谱角函数（ＬSAD）代替均方误差（ＬMSE）作为目标函数。 SSAE法通过梯度下降方式对目标函数进行优化求解获取神经网络参数。 然后, 利用Hapke模型建立不同矿物组合和不同质量分数的两个模拟数据集, 数据集共包括高岭石、 叶腊石、 蒙脱石、 绿泥石、 白云母、 方解石、 赤铁矿、 白云石、 钾长石和褐铁矿10种常见矿物光谱。 最后, 利用SSAE方法对模拟数据集进行端元提取测试, 测试结果与网络结构改进过程中产生的6种情况以及顶点成分分析法（VCA）和基于最小体积的变元切分增量拉格朗日单形体识别算法（SISAL）提取结果进行比较。 实验证明, 本研究提供的是一种盲端元识别方法, 改进后的SSAE神经网络端元提取精度比未完成改进前有明显提升。 SSAE法可以成功识别两个模拟数据集所有的端元, 光谱角距离（spectral angle distance, SAD）的平均误差分别为0.0597和0.0344, 与VCA法提取精度差异较小, 均优于SISAL法的识别结果。 SSAE法为矿物高光谱解混提供了新的方向, 对高光谱遥感的地质应用和高光谱遥感定量分析研究具有较好的促进作用。

Rocks in nature are usually aggregates of various minerals. Due to the low spatial resolution of hyperspectral sensors, the hyperspectral data obtained are mostly the mixing spectrum of mineral components. Affected by factors such as noise interference and intimate mixing characteristics of minerals, endmember extracting and quantifying analysis of minerals is still a hotspot and difficulty topicin current research. Based on the deep learning theory, this study improves the autoencoder structure and proposes a new stacked sparse autoencoders method (SSAE), which provides a new idea for mineral hyperspectral unmixing. First of all, according to the characteristics of mineral mixing spectrum, three improvements have been made: first, the bias term of autoencoder neural network is removed; second, the batch normalization (BN) layer is added in front of the activation function of each hidden layer, and the Relu activation function is used for the final output layer; third, spectral angular distance (LSAD) is used as the objective function instead of the mean square error (ＬMSE). The proposed model obtained the parameters by optimizing the objective function through gradient descent method. Then, two simulation datasets with different mineral combinations and different mass fractions are established by using the Hapke model. The datasets include ten pure minerals, kaolinite, pyrophyllite, montmorillonite, chlorite, muscovite, calcite, hematite, dolomite, potassium feldspar and limonite. Finally, SSAE method is used to test the datasets. Test results of SSAE are compared with the results of six cases in the process of autoencoder network improvement as well as the results of VCA and SISAL. Experiments show that the accuracy of SSAE endmember extraction is greatly improved than before. The SSAE method can successfully identify all endmembers of two data sets. The mean errors of Spectral Angle Distance(SAD) are respectively 0.059 7 and 0.034 4, which is less different from the result of the VCA and is better than there sult of SISAL. SSAE method provides a new ideal for hyperspectral unmixing, and has a better promoting effect on the geological application and quantitative analysis of hyperspectral remote sensing.