高光谱遥感技术能够更细致地检测出岩矿的光谱特征, 为高光谱岩矿识别提供了强有力手段。 基于特定吸收特征波段的高光谱岩矿识别模型依赖很高的先验知识且难以满足区分不同类型岩石的要求, 因此探索建立准确、 高效的高光谱岩石自动识别模型具有重要意义。 在禄丰恐龙谷地区采集三类典型的沉积岩(泥岩、 砂岩和灰岩各21个)作为目标样本, 采用ASD FieldSpec3地物光谱仪获取沉积岩样本在350~2 500 nm范围内的高光谱数据, 对原始光谱进行一阶微分、 连续统去除变换并分析其光谱特征, 采用连续投影(SPA)、 竞争性自适应重加权采样(CARS)和迭代保留信息变量法(IRIV)三种特征变量选择算法选取原始光谱及其变换光谱中的特征波长, 基于全波段和特征波长数据分别建立支持向量机(SVM)和随机森林(RF)识别模型。 结果表明: 三种特征变量选择算法对高光谱数据都具有较好的降维效果, 从原始光谱及两种变换光谱选取出的特征波长数量在7~59个之间。 综合光谱变换处理与特征变量选择算法进行模型测试对比试验, 发现组合连续统去除-SPA-SVM模型方法在识别三类目标沉积岩上的表现最好, 其识别精度为0.952 4, 此时选取出用于输入模型的特征波长数量为10个, 只占全波段的0.5%, 大大降低了模型的运算量, 其中2个特征波长位于550 nm附近的Fe2+和Fe3+吸收带, 2个位于900 nm附近Fe3+吸收带, 5个位于1 900和2 200 nm附近的水分子、 羟基吸收带, 其分布可以较好地反映沉积岩化学成分差异导致的光谱吸收特征规律。 实验结果表明采用光谱变换与特征变量选择算法进行高光谱沉积岩自动识别是可行的, 能为高光谱岩矿识别方法提供参考。

Hyperspectral remote sensing technology can show the spectral characteristics of rocks and minerals in more detail, which provides a powerful means for hyperspectral rock and mineral identification. The traditional hyperspectral rock and mineral identification model based on specific absorption characteristic band depends on high a priori knowledge and is difficult to meet the requirements of distinguishing different types of rocks. Therefore, exploring and establishing an accurate and efficient hyperspectral rock automatic identification model is of great significance. Three typical sedimentary rocks (21 mudstone, sandstone and limestone) were collected as target samples in the Lufeng Dinosaur Valley area. The hyperspectral data of sedimentary rock samples in the range of 350~2 500 nm were obtained with the aid of the ASD fieldspec3 ground feature spectrometer. The original spectrums first-order differential and continuous removal transformation were carried out, and the spectral characteristics were analyzed. The continuous projection algorithm (SPA) was used. Competitive adaptive reweighted sampling algorithm (CARS) and iterative retained information variable method (IRIV) select the characteristic wavelengths in the original spectrum and transformed spectrum and then establish support vector machine (SVM) and random forest (RF) recognition models based on the full band and characteristic wavelength data respectively. The results show that the three feature variable selection algorithms have a good dimensionality reduction effect on hyperspectral data, and the number of feature wavelengths selected from the original spectrum and the two transform spectra is between 7~59. It is obtained that the combined continuum removal SPA-SVM model method performs best for identifying three types of target sedimentary rocks, and its recognition accuracy is 0.952 4. At this time, 10 characteristic wavelengths are selected for the input model, which accounts for only 0.5% of the whole band, which greatly reduces the amount of calculation of the model. Two characteristic wavelengths are located in the Fe2+ and Fe3+ absorption bands near 550 nm, Two Fe3+ absorption bands near 900nm and five water molecules and hydroxyl absorption bands near 1 900, and 2 200 nm can better reflect the spectral absorption characteristics caused by the difference of chemical composition of sedimentary rocks. The experimental results show that the automatic recognition of hyperspectral sedimentary rocks using spectral transformation and characteristic variable selection algorithm is feasible and can provide a reference for hyperspectral rock and mineral recognition methods.