干涉光谱结合偏最小二乘法反演热液CH4的研究

热液释放的高温甲烷气体经扩散作用先后进入海洋和大气，并对地球物理、化学和生物方面产生深刻影响。由于海洋溶解甲烷数据的缺乏，导致人们对深海热液释放甲烷的活动机制和环境效应还缺乏足够的认识。我们前期提出一种光学被动成像干涉系统OPIIS用于热液甲烷浓度、温度和压强的实时探测和长期观测。为了从OPIIS的干涉光谱中精确、稳定、快速的获取热液甲烷信息，采用将干涉光谱与偏最小二乘法相结合的方法处理OPIIS数据。首先分别建立三个甲烷浓度、温度和压强的单因变量预测模型，再利用干涉条纹与辐射光谱的关系，间接建立干涉光谱与甲烷浓度、温度和压强的PLS预测模型，提高了预测模型在实际应用中的抗干扰能力和稳定性。基于洛仑兹线型建立了不同于大气环境的深海气体辐射模型，并利用HITRAN2016分子光谱数据库的光谱参数，建立了深海甲烷在任意浓度、任意温度和任意压强下的辐射光谱数据库。挑选热液其他气体对甲烷探测干扰较小的甲烷泛频带1.64～1.66 μm内的六条谱线建立甲烷辐射光谱与浓度、温度和压强的偏最小二乘回归模型。另外，分析了训练集取样个数、取样间隔和主成分个数对提高预测模型综合性能的作用。利用不同训练集样本数，不同训练集取样间隔和不同的主成分数，分别建立96个浓度、温度和压强预测模型，并分别利用25组预测集样本对预测模型进行交叉验证。不同模型预测均方根误差和决定系数的对比表明，训练集取样个数、取样间隔和主成分个数等单一因素的改变并不能同时提高预测模型的预测精度、稳定性、适用范围和运算量等综合性能。经过平衡选取各项指标确定的最优回归模型的参数为：浓度、温度和压强的适用范围分别为5～375 mmol·L-1，580～678 K，10～34.5 MPa, 浓度、温度和压强的训练集取样个数分别为50组，25组，25组，采样间隔分别为5 mmol·L-1，2 K，0.5 MPa, 浓度、温度和压强预测模型的主成分数分别为2，2，5。浓度、温度和压强预测模型的预测均方根误差分别为3.082×10-6，0.977 0，5.052×10-3，决定系数分别为0.999 9，0.998 9，0.999 9。浓度、温度和压强的预测误差分别为±1.21×10-7，±3.63×10-3，±9.49×10-4，对应的预测精度分别为±45.4 nmol·L-1，±2.5 K，±3.3×10-2 MPa。结果表明，干涉光谱结合偏最小二乘法的反演算法可以精确、稳定、快速的获取热液甲烷气体的浓度、温度和压强信息。

Retrieval of Hydrothermal CH4 Based on Interference Spectroscopy and PLS Methods

The methane (CH₄) gas released by hydrothermal enters into the ocean and atmosphere successively by diffusing and causes inestimable effect on earth in physics, chemistry and biology. The principle and environment effect of abyssal hydrothermal still require further study because limited information is available about dissolved methane. In our previous work, we propose an optical passive imaging interference system (OPIIS) for the real-time detection and long-term observation of hydrothermal methane’s concentration, temperature, and pressure. To accurately, stably, and rapidly obtain the information of hydrothermal methane from OPIIS’s interferogram, this paper processes OPIIS’s data by combining interference spectra and partial least squares (PLS) algorithm. We built three single-dependent variable models between methane radiance spectra and gas concentration, temperature and pressure, respectively. Then we can establish the PLS prediction model between interference fringes indirectly on the basis of relationship between interference fringes and radiance spectra, which can improve the capacity of resisting disturbance and stability of prediction models in practical application. On the basis of Lorentz profile, we build the deep ocean gas emission model different from atmosphere emission and obtain the synthetic methane radiance spectrum database at any concentration, temperature and pressure by using the methane spectral parameters from HITRAN2016 molecular spectroscopy database. The six spectral lines of methane in the range of 1.64～1.66 μm are selected for the PLS regression model between methane radiance spectra and gas concentration, temperature and pressure. Furthermore, this paper analyzes the contribution of number of training samples, interval of training samples and number of principal components to the improvement of the comprehensive performance of regression model. The 96 groups of concentration, temperature and pressure regression model are built by using different groups, intervals and principal components, and those regression models are cross-validated using 25 groups of prediction samples. The comparison results of those regression models’ root mean square error of prediction (RMSEP) and coefficient of determination (R2) indicate that the change of single factors such as the number of training samples, the interval of training samples and the number of principal components can not improve the prediction model’s comprehensive performance about prediction accuracy, stability, application scope and computation. Finally, the optimized model with balanced performance is determined with concentration, temperature and pressure application ranges at 5～375 mmol·L-1，580～678 K，10～34.5 MPa, training samples of concentration, temperature and pressure are 50 groups, 25 groups, 25 groups, intervals at 5 mmol·L-1，2 K，0.5 MPa, principal components are 2，2，5. The RMSEPs of concentration, temperature and pressure are 3.082×10-6，0.977 0，5.052×10-3, and R2s are 0.999 9，0.998 9，0.999 9, respectively. The prediction errors of concentration, temperature and pressure are ±1.21×10-7，±3.63×10-3，±9.49×10-4, and the corresponding precisions are ±45.4 nmol·L-1，±2.5 K，±3.3×10-2 MPa. The results indicate that this retrieval algorithm can accurately, stably, and rapidly obtain concentration, temperature and pressure of hydrothermal methane.