中国科学院遥感与数字地球研究所-高光谱遥感应用技术研究室

【作者】              张学文                                                                                                                                           

【导师】              童庆禧；王晋年；张立福

【 学位年度 】    2012

【论文级别】      博士    

【关键词】          发射率，矿物识别，最小二乘，地面热红外成像光谱系统 

【Key words】   Emissivity; Mineral Determinaton; Least Squares; FISS-TR  

【中文摘要】   

热红外成像光谱数据具有极其广泛的应用需求，特别是在地质领域具有不可替代的作用。目前国际上，美国、日本、加拿大等发达国家已开展了热红外成像光谱的应用研究。相对于星载/机载平台，地面热红外成像光谱仪大大拓宽了其应用领域，其仪器研制正方兴未艾。在可见光近红外波段，我国第一套地面成像光谱辐射测量系统（Filed Imaging Spectrometer System，FISS）研制成功并取得了诸多应用成果。为了进一步扩展FISS的应用范围，使之能够应用于地质领域，推动我国热红外成像光谱技术的进步，地面热红外成像光谱系统（Filed Imaging Spectrometer System –Thermal Infrared，FISS-TR）的研发被列入研究计划。

        本文主要面向FISS-TR，利用国内外已有岩石和矿物的热红外光谱库数据，开展了岩石组分及含量的精细识别模型研究，为FISS-TR的应用提供新方法，同时为FISS-TR成像光谱仪器的技术指标设计提供重要的技术支持。论文的主要工作及成果总结如下：

（1）介绍了矿物发射光谱分类的原理，在包络线去除的基础上，定量研究了碳酸盐、硫酸盐和硅酸盐类矿物在热红外波段的发射光谱特征，包括吸收位置、吸收深度、吸收宽度、吸收面积和吸收对称性。碳酸盐类矿物在热红外波段仅有一个很窄在强吸收谷，在11.3μm左右。硫酸盐类矿物在8.5-9μm的区间有一个明显的宽吸收带；硅酸盐类矿物的发射光谱特征比较复杂，不同结构的硅酸盐矿物发射光谱略有差异，总体表现为复杂的Si-O基团的振动，在8.5～12.0μm区间有强的吸收特征，一般具有多个吸收峰。分析结果表明，同一类的矿物的热红外波段的发射光谱基本上是相似的，通过矿物的发射率光谱特征，可以对岩矿进行定性识别。

（2）在岩石矿物组分精细识别算法研究方面，比较了最小二乘不同约束模型算法反演精度，提出了基于阈值的最小二乘线性解混算法（Threshold Constrained Least Squares，TCLS），从而可以更好地反演出岩石中的矿物端元及其含量。在确定端元的情况下，TCLS模型、无约束模型（Unconstrained Least Squares，ULS）和非负约束模型（Abundance Nonnegativty Constraint，ANC）具有相同的矿物丰度反演结果，其反演精度比和1约束模型（Abundance Sum-to-one Constraint，ASC）和全约束模型（Fully Constrained Least Squares，FCLS）要高，这可能是因为在实际情况中，岩石中的矿物端元并不完备，强行将解得的矿物端元丰度和约束到1，反而会降低解的精度及可靠性。在未知端元的情况下，TCLS模型能够剔除微量矿物对端元识别的影响，具有最优的矿物端元识别能力及丰度反演精度。在一定程度上，ULS模型、ANC模型ASC模型和FSLC模型也能识别矿物端元，但其反演的端元数远多于岩石中实际含有的矿物，降低了其丰度反演的精度。

（3）结合实测发射光谱和数值模拟发射光谱，研究了光谱分辨率、信噪比和包络线去除等因素对发射光谱识别矿物精度的影响。随着光谱分辨率的降低，矿物含量反演精度呈阶梯性降低，其中0.177μm的光谱分辨率是精度变化的重要分水岭。研究结果表明，成像光谱仪器的最佳光谱分辨率约为0.177μm，波段数量最佳为32个，为FISS-TR传感器的设计提供了重要的理论参考。随着信噪比的降低，TCLS模型反演的模型误差逐步增大。在已知端元的条件下，信噪比对矿物丰度反演精度影响较小，当信噪比为30时，反演丰度最大误差<10%，模型误差<5%，；在盲端元的条件下，信噪比对端元识别及其丰度反演有较大影响，当信噪比降低到40时，岩石中含量较低的端元会难以识别。研究结果表明，光谱数据的信噪比至少要达到50，才能够满足模型误差小于1%的精度的要求，为FISS-TR的信噪比参数设计提供重要参考依据。在确定端元的情况下，包络线去除算法可以用于矿物丰度反演。如果对岩石光谱经过包络线去除处理，则矿物端元光谱同样要去除包络线，其丰度反演精度最大误差小于2%，模型误差在1%左右。

 【Abstract】 

Thermal infrared imaging spectrometer has broad applications. Presently, several developed countries carried out the application study of thermal infrared imaging. Compared to the satellite / airborne platforms, field thermal infrared imaging spectrometer has more broaden applications, and its instrument development is in the ascendant.The first self-developed Filed Imaging Spectrometer System (FISS) is successfully developed in the visible/near-infrared region and gets some good application results. In order to further expand the application range of the FISS, to be applied to the geology, the plan about thermal infrared (FISS-TR) region is going to be developed.

      To promote the FISS-TR on the development of rock and mineral identification, as well as to provide technical parameters of the instrument manufacturer, the dissertation carries out basic research on mineral end member and abundance determination using thermal infrared hyperspectral data. Main work and results are summarized as follows:

1. The mineral emission spectra classification principle is introduced. Based on the continuum removed, the characteristics of the emission spectra, including absorption position, absorption depth, absorption width, absorption area, and absorption symmetry, of the carbonate, sulphate and silicate minerals are quantitative studied in the thermal infrared bands. Carbonate minerals have only a very narrow, strong absorption valley, around 11.3μmin the thermal infrared band. Sulfate minerals have an obvious broad absorption features in the range 8.5～9μm.The emission spectra features of the silicate minerals are more complex than others, which different structures of silicate minerals could display slightly different emission spectra features. Because of complex Si-O vibration, silicate minerals show strong absorption features in the 8.5 ~ 12.0μminterval and have multiple absorption peaks. The results show that the emission spectra of the same type of minerals in thermal infrared bands are basically similar. It is possible to identify the rock and mineral by analysis of mineral emissivity spectral features.

2. The dissertation studies the algorithm for recognition rock and determination of mineral composition, compares mineral inversion accuracy of the least squares algorithm using different constraint model, and presents the threshold constrained least squares linear unmixing algorithm (TCLS) which could retrieve mineral end member and their abundance better. Under the condition of known end member mineral, TCLS model, ULS model and ANC model could obtain the same mineral abundance, the retrieval accuracy better than ASC model and FCLS model. This may be due to the fact that end member of rocks are not complete in reality. The abundance sum-to-one constrainet might reduce the accuracy and reliability of the results. In the case of the unknown end member, TCLS model can remove trace mineral end member obstruction, identify end member mineral and get optimal abundance inversion accuracy. ULS model, ANC model, ASC model and the FSLC model can also identify the mineral end member, but inversion end member are far more than the mineral contained in rock, as well as reduce the accuracy of the mineral abundance.

3. With combination of the measured spectra and numerical simulation spectra, the impact of the spectral resolution, signal-to-noise ratio (SNR) and continuum removal on determination of mineral in thermal infrared band have been studied, respectively. The lower spectral resolution is, the worse retrieval results of mineral end member and abundance are obtained. In the case of unknown end members，TCLS model can correctly identify the mineral end member and its model difference vary small with spectral resolution declining from 279 bands (0.021μm) to 32 bands (0.177μm). When the spectral resolution is less than to 32 bands, model difference increases fast, and results of identification of the mineral become incorrect. So, 32 bands can be taken as reference for design of FISS-TR sensor. With the SNR declining, the model retrieval error becomes higher. When SNR reaches 30, the maximum error of the abundance is less than 10%, model error is less than 5%. If the end member is unknown, SNR impacts on inversion accuracy more seriously. While SNR equals 40, it is difficult to identify the mineral end member.The results show that in order to meet the model error less than 1%, the SNR of the data could achieve 50 at least.It is important for FISS-TR to design SNR parameter. In the case of known end member, the continuum removal algorithm can be used for mineral abundance inversion. If rock spectrum is preprocessed by continuum removal, the mineral spectrum also needs to be. The maximum error of abundance is less than 2%; the model error is about 1%.